2 * Copyright 2000, International Business Machines Corporation and others.
5 * This software has been released under the terms of the IBM Public
6 * License. For details, see the LICENSE file in the top-level source
7 * directory or online at http://www.openafs.org/dl/license10.html
13 #include <afsconfig.h>
14 #include "afs/param.h"
17 #include "afs/sysincludes.h" /*Standard vendor system headers */
18 #include "afsincludes.h" /*AFS-based standard headers */
19 #include "afs/afs_stats.h" /* statistics */
20 #include "afs/afs_cbqueue.h"
21 #include "afs/afs_osidnlc.h"
23 /* Forward declarations. */
24 static void afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint);
25 static int afs_FreeDiscardedDCache(void);
26 static void afs_DiscardDCache(struct dcache *);
27 static void afs_FreeDCache(struct dcache *);
29 static afs_int32 afs_DCGetBucket(struct vcache *);
30 static void afs_DCAdjustSize(struct dcache *, afs_int32, afs_int32);
31 static void afs_DCMoveBucket(struct dcache *, afs_int32, afs_int32);
32 static void afs_DCSizeInit(void);
33 static afs_int32 afs_DCWhichBucket(afs_int32, afs_int32);
36 * --------------------- Exported definitions ---------------------
39 afs_int32 afs_blocksUsed_0; /*!< 1K blocks in cache - in theory is zero */
40 afs_int32 afs_blocksUsed_1; /*!< 1K blocks in cache */
41 afs_int32 afs_blocksUsed_2; /*!< 1K blocks in cache */
42 afs_int32 afs_pct1 = -1;
43 afs_int32 afs_pct2 = -1;
44 afs_uint32 afs_tpct1 = 0;
45 afs_uint32 afs_tpct2 = 0;
46 afs_uint32 splitdcache = 0;
48 afs_lock_t afs_xdcache; /*!< Lock: alloc new disk cache entries */
49 afs_int32 afs_freeDCList; /*!< Free list for disk cache entries */
50 afs_int32 afs_freeDCCount; /*!< Count of elts in freeDCList */
51 afs_int32 afs_discardDCList; /*!< Discarded disk cache entries */
52 afs_int32 afs_discardDCCount; /*!< Count of elts in discardDCList */
53 struct dcache *afs_freeDSList; /*!< Free list for disk slots */
54 struct dcache *afs_Initial_freeDSList; /*!< Initial list for above */
55 afs_dcache_id_t cacheInode; /*!< Inode for CacheItems file */
56 struct osi_file *afs_cacheInodep = 0; /*!< file for CacheItems inode */
57 struct afs_q afs_DLRU; /*!< dcache LRU */
58 afs_int32 afs_dhashsize = 1024;
59 afs_int32 *afs_dvhashTbl; /*!< Data cache hash table: hashed by FID + chunk number. */
60 afs_int32 *afs_dchashTbl; /*!< Data cache hash table: hashed by FID. */
61 afs_int32 *afs_dvnextTbl; /*!< Dcache hash table links */
62 afs_int32 *afs_dcnextTbl; /*!< Dcache hash table links */
63 struct dcache **afs_indexTable; /*!< Pointers to dcache entries */
64 afs_hyper_t *afs_indexTimes; /*!< Dcache entry Access times */
65 afs_int32 *afs_indexUnique; /*!< dcache entry Fid.Unique */
66 unsigned char *afs_indexFlags; /*!< (only one) Is there data there? */
67 afs_hyper_t afs_indexCounter; /*!< Fake time for marking index
69 afs_int32 afs_cacheFiles = 0; /*!< Size of afs_indexTable */
70 afs_int32 afs_cacheBlocks; /*!< 1K blocks in cache */
71 afs_int32 afs_cacheStats; /*!< Stat entries in cache */
72 afs_int32 afs_blocksUsed; /*!< Number of blocks in use */
73 afs_int32 afs_blocksDiscarded; /*!<Blocks freed but not truncated */
74 afs_int32 afs_fsfragsize = AFS_MIN_FRAGSIZE; /*!< Underlying Filesystem minimum unit
75 *of disk allocation usually 1K
76 *this value is (truefrag -1 ) to
77 *save a bunch of subtracts... */
78 #ifdef AFS_64BIT_CLIENT
79 #ifdef AFS_VM_RDWR_ENV
80 afs_size_t afs_vmMappingEnd; /* !< For large files (>= 2GB) the VM
81 * mapping an 32bit addressing machines
82 * can only be used below the 2 GB
83 * line. From this point upwards we
84 * must do direct I/O into the cache
85 * files. The value should be on a
87 #endif /* AFS_VM_RDWR_ENV */
88 #endif /* AFS_64BIT_CLIENT */
90 /* The following is used to ensure that new dcache's aren't obtained when
91 * the cache is nearly full.
93 int afs_WaitForCacheDrain = 0;
94 int afs_TruncateDaemonRunning = 0;
95 int afs_CacheTooFull = 0;
97 afs_int32 afs_dcentries; /*!< In-memory dcache entries */
100 int dcacheDisabled = 0;
102 struct afs_cacheOps afs_UfsCacheOps = {
103 #ifndef HAVE_STRUCT_LABEL_SUPPORT
116 .truncate = osi_UFSTruncate,
117 .fread = afs_osi_Read,
118 .fwrite = afs_osi_Write,
119 .close = osi_UFSClose,
120 .vreadUIO = afs_UFSReadUIO,
121 .vwriteUIO = afs_UFSWriteUIO,
122 .GetDSlot = afs_UFSGetDSlot,
123 .GetVolSlot = afs_UFSGetVolSlot,
124 .HandleLink = afs_UFSHandleLink,
128 struct afs_cacheOps afs_MemCacheOps = {
129 #ifndef HAVE_STRUCT_LABEL_SUPPORT
131 afs_MemCacheTruncate,
141 .open = afs_MemCacheOpen,
142 .truncate = afs_MemCacheTruncate,
143 .fread = afs_MemReadBlk,
144 .fwrite = afs_MemWriteBlk,
145 .close = afs_MemCacheClose,
146 .vreadUIO = afs_MemReadUIO,
147 .vwriteUIO = afs_MemWriteUIO,
148 .GetDSlot = afs_MemGetDSlot,
149 .GetVolSlot = afs_MemGetVolSlot,
150 .HandleLink = afs_MemHandleLink,
154 int cacheDiskType; /*Type of backing disk for cache */
155 struct afs_cacheOps *afs_cacheType;
158 * Where is this vcache's entry associated dcache located/
159 * \param avc The vcache entry.
160 * \return Bucket index:
165 afs_DCGetBucket(struct vcache *avc)
170 /* This should be replaced with some sort of user configurable function */
171 if (avc->f.states & CRO) {
173 } else if (avc->f.states & CBackup) {
183 * Readjust a dcache's size.
185 * \param adc The dcache to be adjusted.
186 * \param oldSize Old size for the dcache.
187 * \param newSize The new size to be adjusted to.
191 afs_DCAdjustSize(struct dcache *adc, afs_int32 oldSize, afs_int32 newSize)
193 afs_int32 adjustSize = newSize - oldSize;
201 afs_blocksUsed_0 += adjustSize;
202 afs_stats_cmperf.cacheBucket0_Discarded += oldSize;
205 afs_blocksUsed_1 += adjustSize;
206 afs_stats_cmperf.cacheBucket1_Discarded += oldSize;
209 afs_blocksUsed_2 += adjustSize;
210 afs_stats_cmperf.cacheBucket2_Discarded += oldSize;
218 * Move a dcache from one bucket to another.
220 * \param adc Operate on this dcache.
221 * \param size Size in bucket (?).
222 * \param newBucket Destination bucket.
226 afs_DCMoveBucket(struct dcache *adc, afs_int32 size, afs_int32 newBucket)
231 /* Substract size from old bucket. */
235 afs_blocksUsed_0 -= size;
238 afs_blocksUsed_1 -= size;
241 afs_blocksUsed_2 -= size;
245 /* Set new bucket and increase destination bucket size. */
246 adc->bucket = newBucket;
251 afs_blocksUsed_0 += size;
254 afs_blocksUsed_1 += size;
257 afs_blocksUsed_2 += size;
265 * Init split caches size.
270 afs_blocksUsed_0 = afs_blocksUsed_1 = afs_blocksUsed_2 = 0;
279 afs_DCWhichBucket(afs_int32 phase, afs_int32 bucket)
284 afs_pct1 = afs_blocksUsed_1 / (afs_cacheBlocks / 100);
285 afs_pct2 = afs_blocksUsed_2 / (afs_cacheBlocks / 100);
287 /* Short cut: if we don't know about it, try to kill it */
288 if (phase < 2 && afs_blocksUsed_0)
291 if (afs_pct1 > afs_tpct1)
293 if (afs_pct2 > afs_tpct2)
295 return 0; /* unlikely */
300 * Warn about failing to store a file.
302 * \param acode Associated error code.
303 * \param avolume Volume involved.
304 * \param aflags How to handle the output:
305 * aflags & 1: Print out on console
306 * aflags & 2: Print out on controlling tty
308 * \note Environment: Call this from close call when vnodeops is RCS unlocked.
312 afs_StoreWarn(afs_int32 acode, afs_int32 avolume,
315 static char problem_fmt[] =
316 "afs: failed to store file in volume %d (%s)\n";
317 static char problem_fmt_w_error[] =
318 "afs: failed to store file in volume %d (error %d)\n";
319 static char netproblems[] = "network problems";
320 static char partfull[] = "partition full";
321 static char overquota[] = "over quota";
323 AFS_STATCNT(afs_StoreWarn);
329 afs_warn(problem_fmt, avolume, netproblems);
331 afs_warnuser(problem_fmt, avolume, netproblems);
332 } else if (acode == ENOSPC) {
337 afs_warn(problem_fmt, avolume, partfull);
339 afs_warnuser(problem_fmt, avolume, partfull);
342 /* EDQUOT doesn't exist on solaris and won't be sent by the server.
343 * Instead ENOSPC will be sent...
345 if (acode == EDQUOT) {
350 afs_warn(problem_fmt, avolume, overquota);
352 afs_warnuser(problem_fmt, avolume, overquota);
360 afs_warn(problem_fmt_w_error, avolume, acode);
362 afs_warnuser(problem_fmt_w_error, avolume, acode);
367 * Try waking up truncation daemon, if it's worth it.
370 afs_MaybeWakeupTruncateDaemon(void)
372 if (!afs_CacheTooFull && afs_CacheIsTooFull()) {
373 afs_CacheTooFull = 1;
374 if (!afs_TruncateDaemonRunning)
375 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
376 } else if (!afs_TruncateDaemonRunning
377 && afs_blocksDiscarded > CM_MAXDISCARDEDCHUNKS) {
378 afs_osi_Wakeup((int *)afs_CacheTruncateDaemon);
385 * Keep statistics on run time for afs_CacheTruncateDaemon. This is a
386 * struct so we need only export one symbol for AIX.
388 static struct CTD_stats {
389 osi_timeval_t CTD_beforeSleep;
390 osi_timeval_t CTD_afterSleep;
391 osi_timeval_t CTD_sleepTime;
392 osi_timeval_t CTD_runTime;
396 u_int afs_min_cache = 0;
399 * If there are waiters for the cache to drain, wake them if
400 * the number of free cache blocks reaches the CM_CACHESIZEDDRAINEDPCT.
403 * This routine must be called with the afs_xdcache lock held
407 afs_WakeCacheWaitersIfDrained(void)
409 if (afs_WaitForCacheDrain) {
410 if ((afs_blocksUsed - afs_blocksDiscarded) <=
411 PERCENT(CM_CACHESIZEDRAINEDPCT, afs_cacheBlocks)) {
412 afs_WaitForCacheDrain = 0;
413 afs_osi_Wakeup(&afs_WaitForCacheDrain);
419 * Keeps the cache clean and free by truncating uneeded files, when used.
424 afs_CacheTruncateDaemon(void)
426 osi_timeval_t CTD_tmpTime;
430 PERCENT((100 - CM_DCACHECOUNTFREEPCT + CM_DCACHEEXTRAPCT), afs_cacheFiles);
432 (((10 * AFS_CHUNKSIZE(0)) + afs_fsfragsize) & ~afs_fsfragsize) >> 10;
434 osi_GetuTime(&CTD_stats.CTD_afterSleep);
435 afs_TruncateDaemonRunning = 1;
437 cb_lowat = PERCENT((CM_DCACHESPACEFREEPCT - CM_DCACHEEXTRAPCT), afs_cacheBlocks);
438 ObtainWriteLock(&afs_xdcache, 266);
439 if (afs_CacheTooFull || afs_WaitForCacheDrain) {
440 int space_needed, slots_needed;
441 /* if we get woken up, we should try to clean something out */
442 for (counter = 0; counter < 10; counter++) {
444 afs_blocksUsed - afs_blocksDiscarded - cb_lowat;
445 if (space_needed < 0)
448 dc_hiwat - afs_freeDCCount - afs_discardDCCount;
449 if (slots_needed < 0)
451 if (slots_needed || space_needed)
452 afs_GetDownD(slots_needed, &space_needed, 0);
453 if ((space_needed <= 0) && (slots_needed <= 0)) {
454 afs_CacheTooFull = 0;
457 if (afs_termState == AFSOP_STOP_TRUNCDAEMON)
460 if (!afs_CacheIsTooFull()) {
461 afs_CacheTooFull = 0;
462 afs_WakeCacheWaitersIfDrained();
464 } /* end of cache cleanup */
465 ReleaseWriteLock(&afs_xdcache);
468 * This is a defensive check to try to avoid starving threads
469 * that may need the global lock so thay can help free some
470 * cache space. If this thread won't be sleeping or truncating
471 * any cache files then give up the global lock so other
472 * threads get a chance to run.
474 if ((afs_termState != AFSOP_STOP_TRUNCDAEMON) && afs_CacheTooFull
475 && (!afs_blocksDiscarded || afs_WaitForCacheDrain)) {
476 afs_osi_Wait(100, 0, 0); /* 100 milliseconds */
480 * This is where we free the discarded cache elements.
482 while (afs_blocksDiscarded && !afs_WaitForCacheDrain
483 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
484 int code = afs_FreeDiscardedDCache();
486 /* If we can't free any discarded dcache entries, that's okay.
487 * We're just doing this in the background; if someone needs
488 * discarded entries freed, they will try it themselves and/or
489 * signal us that the cache is too full. In any case, we'll
490 * try doing this again the next time we run through the loop.
496 /* See if we need to continue to run. Someone may have
497 * signalled us while we were executing.
499 if (!afs_WaitForCacheDrain && !afs_CacheTooFull
500 && (afs_termState != AFSOP_STOP_TRUNCDAEMON)) {
501 /* Collect statistics on truncate daemon. */
502 CTD_stats.CTD_nSleeps++;
503 osi_GetuTime(&CTD_stats.CTD_beforeSleep);
504 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_afterSleep,
505 CTD_stats.CTD_beforeSleep);
506 afs_stats_AddTo(CTD_stats.CTD_runTime, CTD_tmpTime);
508 afs_TruncateDaemonRunning = 0;
509 afs_osi_Sleep((int *)afs_CacheTruncateDaemon);
510 afs_TruncateDaemonRunning = 1;
512 osi_GetuTime(&CTD_stats.CTD_afterSleep);
513 afs_stats_GetDiff(CTD_tmpTime, CTD_stats.CTD_beforeSleep,
514 CTD_stats.CTD_afterSleep);
515 afs_stats_AddTo(CTD_stats.CTD_sleepTime, CTD_tmpTime);
517 if (afs_termState == AFSOP_STOP_TRUNCDAEMON) {
518 afs_termState = AFSOP_STOP_AFSDB;
519 afs_osi_Wakeup(&afs_termState);
527 * Make adjustment for the new size in the disk cache entry
529 * \note Major Assumptions Here:
530 * Assumes that frag size is an integral power of two, less one,
531 * and that this is a two's complement machine. I don't
532 * know of any filesystems which violate this assumption...
534 * \param adc Ptr to dcache entry.
535 * \param anewsize New size desired.
540 afs_AdjustSize(struct dcache *adc, afs_int32 newSize)
544 AFS_STATCNT(afs_AdjustSize);
546 adc->dflags |= DFEntryMod;
547 oldSize = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
548 adc->f.chunkBytes = newSize;
551 newSize = ((newSize + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
552 afs_DCAdjustSize(adc, oldSize, newSize);
553 if ((newSize > oldSize) && !AFS_IS_DISCONNECTED) {
555 /* We're growing the file, wakeup the daemon */
556 afs_MaybeWakeupTruncateDaemon();
558 afs_blocksUsed += (newSize - oldSize);
559 afs_stats_cmperf.cacheBlocksInUse = afs_blocksUsed; /* XXX */
564 * This routine is responsible for moving at least one entry (but up
565 * to some number of them) from the LRU queue to the free queue.
567 * \param anumber Number of entries that should ideally be moved.
568 * \param aneedSpace How much space we need (1K blocks);
571 * The anumber parameter is just a hint; at least one entry MUST be
572 * moved, or we'll panic. We must be called with afs_xdcache
573 * write-locked. We should try to satisfy both anumber and aneedspace,
574 * whichever is more demanding - need to do several things:
575 * 1. only grab up to anumber victims if aneedSpace <= 0, not
576 * the whole set of MAXATONCE.
577 * 2. dynamically choose MAXATONCE to reflect severity of
578 * demand: something like (*aneedSpace >> (logChunk - 9))
580 * \note N.B. if we're called with aneedSpace <= 0 and anumber > 0, that
581 * indicates that the cache is not properly configured/tuned or
582 * something. We should be able to automatically correct that problem.
585 #define MAXATONCE 16 /* max we can obtain at once */
587 afs_GetDownD(int anumber, int *aneedSpace, afs_int32 buckethint)
591 struct VenusFid *afid;
596 afs_uint32 victims[MAXATONCE];
597 struct dcache *victimDCs[MAXATONCE];
598 afs_hyper_t victimTimes[MAXATONCE]; /* youngest (largest LRU time) first */
599 afs_uint32 victimPtr; /* next free item in victim arrays */
600 afs_hyper_t maxVictimTime; /* youngest (largest LRU time) victim */
601 afs_uint32 maxVictimPtr; /* where it is */
605 AFS_STATCNT(afs_GetDownD);
607 if (CheckLock(&afs_xdcache) != -1)
608 osi_Panic("getdownd nolock");
609 /* decrement anumber first for all dudes in free list */
610 /* SHOULD always decrement anumber first, even if aneedSpace >0,
611 * because we should try to free space even if anumber <=0 */
612 if (!aneedSpace || *aneedSpace <= 0) {
613 anumber -= afs_freeDCCount;
615 return; /* enough already free */
619 /* bounds check parameter */
620 if (anumber > MAXATONCE)
621 anumber = MAXATONCE; /* all we can do */
623 /* rewrite so phases include a better eligiblity for gc test*/
625 * The phase variable manages reclaims. Set to 0, the first pass,
626 * we don't reclaim active entries, or other than target bucket.
627 * Set to 1, we reclaim even active ones in target bucket.
628 * Set to 2, we reclaim any inactive one.
629 * Set to 3, we reclaim even active ones. On Solaris, we also reclaim
630 * entries whose corresponding vcache has a nonempty multiPage list, when
639 for (i = 0; i < afs_cacheFiles; i++)
640 /* turn off all flags */
641 afs_indexFlags[i] &= ~IFFlag;
643 while (anumber > 0 || (aneedSpace && *aneedSpace > 0)) {
644 /* find oldest entries for reclamation */
645 maxVictimPtr = victimPtr = 0;
646 hzero(maxVictimTime);
647 curbucket = afs_DCWhichBucket(phase, buckethint);
648 /* select victims from access time array */
649 for (i = 0; i < afs_cacheFiles; i++) {
650 if (afs_indexFlags[i] & (IFDataMod | IFFree | IFDiscarded)) {
651 /* skip if dirty or already free */
654 tdc = afs_indexTable[i];
655 if (tdc && (curbucket != tdc->bucket) && (phase < 4))
657 /* Wrong bucket; can't use it! */
660 if (tdc && (tdc->refCount != 0)) {
661 /* Referenced; can't use it! */
664 hset(vtime, afs_indexTimes[i]);
666 /* if we've already looked at this one, skip it */
667 if (afs_indexFlags[i] & IFFlag)
670 if (victimPtr < MAXATONCE) {
671 /* if there's at least one free victim slot left */
672 victims[victimPtr] = i;
673 hset(victimTimes[victimPtr], vtime);
674 if (hcmp(vtime, maxVictimTime) > 0) {
675 hset(maxVictimTime, vtime);
676 maxVictimPtr = victimPtr;
679 } else if (hcmp(vtime, maxVictimTime) < 0) {
681 * We're older than youngest victim, so we replace at
684 /* find youngest (largest LRU) victim */
687 osi_Panic("getdownd local");
689 hset(victimTimes[j], vtime);
690 /* recompute maxVictimTime */
691 hset(maxVictimTime, vtime);
692 for (j = 0; j < victimPtr; j++)
693 if (hcmp(maxVictimTime, victimTimes[j]) < 0) {
694 hset(maxVictimTime, victimTimes[j]);
700 /* now really reclaim the victims */
701 j = 0; /* flag to track if we actually got any of the victims */
702 /* first, hold all the victims, since we're going to release the lock
703 * during the truncate operation.
705 for (i = 0; i < victimPtr; i++) {
706 tdc = afs_GetValidDSlot(victims[i]);
707 /* We got tdc->tlock(R) here */
708 if (tdc && tdc->refCount == 1)
713 ReleaseReadLock(&tdc->tlock);
718 for (i = 0; i < victimPtr; i++) {
719 /* q is first elt in dcache entry */
721 /* now, since we're dropping the afs_xdcache lock below, we
722 * have to verify, before proceeding, that there are no other
723 * references to this dcache entry, even now. Note that we
724 * compare with 1, since we bumped it above when we called
725 * afs_GetValidDSlot to preserve the entry's identity.
727 if (tdc && tdc->refCount == 1) {
728 unsigned char chunkFlags;
729 afs_size_t tchunkoffset = 0;
731 /* xdcache is lower than the xvcache lock */
732 ReleaseWriteLock(&afs_xdcache);
733 ObtainReadLock(&afs_xvcache);
734 tvc = afs_FindVCache(afid, 0, 0 /* no stats, no vlru */ );
735 ReleaseReadLock(&afs_xvcache);
736 ObtainWriteLock(&afs_xdcache, 527);
738 if (tdc->refCount > 1)
741 tchunkoffset = AFS_CHUNKTOBASE(tdc->f.chunk);
742 chunkFlags = afs_indexFlags[tdc->index];
743 if (((phase & 1) == 0) && osi_Active(tvc))
745 if (((phase & 1) == 1) && osi_Active(tvc)
746 && (tvc->f.states & CDCLock)
747 && (chunkFlags & IFAnyPages))
749 if (chunkFlags & IFDataMod)
751 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
752 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, skip,
753 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
754 ICL_HANDLE_OFFSET(tchunkoffset));
756 #if defined(AFS_SUN5_ENV)
758 * Now we try to invalidate pages. We do this only for
759 * Solaris. For other platforms, it's OK to recycle a
760 * dcache entry out from under a page, because the strategy
761 * function can call afs_GetDCache().
763 if (!skip && (chunkFlags & IFAnyPages)) {
766 ReleaseWriteLock(&afs_xdcache);
767 ObtainWriteLock(&tvc->vlock, 543);
768 if (!QEmpty(&tvc->multiPage)) {
769 if (phase < 3 || osi_VM_MultiPageConflict(tvc, tdc)) {
774 /* block locking pages */
775 tvc->vstates |= VPageCleaning;
776 /* block getting new pages */
778 ReleaseWriteLock(&tvc->vlock);
779 /* One last recheck */
780 ObtainWriteLock(&afs_xdcache, 333);
781 chunkFlags = afs_indexFlags[tdc->index];
782 if (tdc->refCount > 1 || (chunkFlags & IFDataMod)
783 || (osi_Active(tvc) && (tvc->f.states & CDCLock)
784 && (chunkFlags & IFAnyPages))) {
786 ReleaseWriteLock(&afs_xdcache);
789 ReleaseWriteLock(&afs_xdcache);
791 code = osi_VM_GetDownD(tvc, tdc);
793 ObtainWriteLock(&afs_xdcache, 269);
794 /* we actually removed all pages, clean and dirty */
796 afs_indexFlags[tdc->index] &=
797 ~(IFDirtyPages | IFAnyPages);
800 ReleaseWriteLock(&afs_xdcache);
802 ObtainWriteLock(&tvc->vlock, 544);
803 if (--tvc->activeV == 0
804 && (tvc->vstates & VRevokeWait)) {
805 tvc->vstates &= ~VRevokeWait;
806 afs_osi_Wakeup((char *)&tvc->vstates);
809 if (tvc->vstates & VPageCleaning) {
810 tvc->vstates &= ~VPageCleaning;
811 afs_osi_Wakeup((char *)&tvc->vstates);
814 ReleaseWriteLock(&tvc->vlock);
816 #endif /* AFS_SUN5_ENV */
818 ReleaseWriteLock(&afs_xdcache);
821 afs_PutVCache(tvc); /*XXX was AFS_FAST_RELE?*/
822 ObtainWriteLock(&afs_xdcache, 528);
823 if (afs_indexFlags[tdc->index] &
824 (IFDataMod | IFDirtyPages | IFAnyPages))
826 if (tdc->refCount > 1)
829 #if defined(AFS_SUN5_ENV)
831 /* no vnode, so IFDirtyPages is spurious (we don't
832 * sweep dcaches on vnode recycling, so we can have
833 * DIRTYPAGES set even when all pages are gone). Just
835 * Hold vcache lock to prevent vnode from being
836 * created while we're clearing IFDirtyPages.
838 afs_indexFlags[tdc->index] &=
839 ~(IFDirtyPages | IFAnyPages);
843 /* skip this guy and mark him as recently used */
844 afs_indexFlags[tdc->index] |= IFFlag;
845 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
846 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 2,
847 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
848 ICL_HANDLE_OFFSET(tchunkoffset));
850 /* flush this dude from the data cache and reclaim;
851 * first, make sure no one will care that we damage
852 * it, by removing it from all hash tables. Then,
853 * melt it down for parts. Note that any concurrent
854 * (new possibility!) calls to GetDownD won't touch
855 * this guy because his reference count is > 0. */
856 afs_Trace4(afs_iclSetp, CM_TRACE_GETDOWND,
857 ICL_TYPE_POINTER, tvc, ICL_TYPE_INT32, 3,
858 ICL_TYPE_INT32, tdc->index, ICL_TYPE_OFFSET,
859 ICL_HANDLE_OFFSET(tchunkoffset));
860 AFS_STATCNT(afs_gget);
861 afs_HashOutDCache(tdc, 1);
862 if (tdc->f.chunkBytes != 0) {
866 (tdc->f.chunkBytes + afs_fsfragsize) >> 10;
871 afs_DiscardDCache(tdc);
876 j = 1; /* we reclaimed at least one victim */
881 } /* end of for victims loop */
884 /* Phase is 0 and no one was found, so try phase 1 (ignore
885 * osi_Active flag) */
888 for (i = 0; i < afs_cacheFiles; i++)
889 /* turn off all flags */
890 afs_indexFlags[i] &= ~IFFlag;
893 /* found no one in phases 0-5, we're hosed */
897 } /* big while loop */
905 * Remove adc from any hash tables that would allow it to be located
906 * again by afs_FindDCache or afs_GetDCache.
908 * \param adc Pointer to dcache entry to remove from hash tables.
910 * \note Locks: Must have the afs_xdcache lock write-locked to call this function.
914 afs_HashOutDCache(struct dcache *adc, int zap)
918 AFS_STATCNT(afs_glink);
920 /* we know this guy's in the LRUQ. We'll move dude into DCQ below */
922 /* if this guy is in the hash table, pull him out */
923 if (adc->f.fid.Fid.Volume != 0) {
924 /* remove entry from first hash chains */
925 i = DCHash(&adc->f.fid, adc->f.chunk);
926 us = afs_dchashTbl[i];
927 if (us == adc->index) {
928 /* first dude in the list */
929 afs_dchashTbl[i] = afs_dcnextTbl[adc->index];
931 /* somewhere on the chain */
932 while (us != NULLIDX) {
933 if (afs_dcnextTbl[us] == adc->index) {
934 /* found item pointing at the one to delete */
935 afs_dcnextTbl[us] = afs_dcnextTbl[adc->index];
938 us = afs_dcnextTbl[us];
941 osi_Panic("dcache hc");
943 /* remove entry from *other* hash chain */
944 i = DVHash(&adc->f.fid);
945 us = afs_dvhashTbl[i];
946 if (us == adc->index) {
947 /* first dude in the list */
948 afs_dvhashTbl[i] = afs_dvnextTbl[adc->index];
950 /* somewhere on the chain */
951 while (us != NULLIDX) {
952 if (afs_dvnextTbl[us] == adc->index) {
953 /* found item pointing at the one to delete */
954 afs_dvnextTbl[us] = afs_dvnextTbl[adc->index];
957 us = afs_dvnextTbl[us];
960 osi_Panic("dcache hv");
965 /* prevent entry from being found on a reboot (it is already out of
966 * the hash table, but after a crash, we just look at fid fields of
967 * stable (old) entries).
969 adc->f.fid.Fid.Volume = 0; /* invalid */
971 /* mark entry as modified */
972 adc->dflags |= DFEntryMod;
977 } /*afs_HashOutDCache */
980 * Flush the given dcache entry, pulling it from hash chains
981 * and truncating the associated cache file.
983 * \param adc Ptr to dcache entry to flush.
986 * This routine must be called with the afs_xdcache lock held
990 afs_FlushDCache(struct dcache *adc)
992 AFS_STATCNT(afs_FlushDCache);
994 * Bump the number of cache files flushed.
996 afs_stats_cmperf.cacheFlushes++;
998 /* remove from all hash tables */
999 afs_HashOutDCache(adc, 1);
1001 /* Free its space; special case null operation, since truncate operation
1002 * in UFS is slow even in this case, and this allows us to pre-truncate
1003 * these files at more convenient times with fewer locks set
1004 * (see afs_GetDownD).
1006 if (adc->f.chunkBytes != 0) {
1007 afs_DiscardDCache(adc);
1008 afs_MaybeWakeupTruncateDaemon();
1010 afs_FreeDCache(adc);
1012 } /*afs_FlushDCache */
1016 * Put a dcache entry on the free dcache entry list.
1018 * \param adc dcache entry to free.
1020 * \note Environment: called with afs_xdcache lock write-locked.
1023 afs_FreeDCache(struct dcache *adc)
1025 /* Thread on free list, update free list count and mark entry as
1026 * freed in its indexFlags element. Also, ensure DCache entry gets
1027 * written out (set DFEntryMod).
1030 afs_dvnextTbl[adc->index] = afs_freeDCList;
1031 afs_freeDCList = adc->index;
1033 afs_indexFlags[adc->index] |= IFFree;
1034 adc->dflags |= DFEntryMod;
1036 afs_WakeCacheWaitersIfDrained();
1037 } /* afs_FreeDCache */
1040 * Discard the cache element by moving it to the discardDCList.
1041 * This puts the cache element into a quasi-freed state, where
1042 * the space may be reused, but the file has not been truncated.
1044 * \note Major Assumptions Here:
1045 * Assumes that frag size is an integral power of two, less one,
1046 * and that this is a two's complement machine. I don't
1047 * know of any filesystems which violate this assumption...
1049 * \param adr Ptr to dcache entry.
1051 * \note Environment:
1052 * Must be called with afs_xdcache write-locked.
1056 afs_DiscardDCache(struct dcache *adc)
1060 AFS_STATCNT(afs_DiscardDCache);
1062 osi_Assert(adc->refCount == 1);
1064 size = ((adc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1065 afs_blocksDiscarded += size;
1066 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1068 afs_dvnextTbl[adc->index] = afs_discardDCList;
1069 afs_discardDCList = adc->index;
1070 afs_discardDCCount++;
1072 adc->f.fid.Fid.Volume = 0;
1073 adc->dflags |= DFEntryMod;
1074 afs_indexFlags[adc->index] |= IFDiscarded;
1076 afs_WakeCacheWaitersIfDrained();
1077 } /*afs_DiscardDCache */
1080 * Get a dcache entry from the discard or free list
1082 * @param[in] indexp A pointer to the head of the dcache free list or discard
1083 * list (afs_freeDCList, or afs_discardDCList)
1085 * @return A dcache from that list, or NULL if none could be retrieved.
1087 * @pre afs_xdcache is write-locked
1089 static struct dcache *
1090 afs_GetDSlotFromList(afs_int32 *indexp)
1094 for ( ; *indexp != NULLIDX; indexp = &afs_dvnextTbl[*indexp]) {
1095 tdc = afs_GetUnusedDSlot(*indexp);
1097 osi_Assert(tdc->refCount == 1);
1098 ReleaseReadLock(&tdc->tlock);
1099 *indexp = afs_dvnextTbl[tdc->index];
1100 afs_dvnextTbl[tdc->index] = NULLIDX;
1108 * Free the next element on the list of discarded cache elements.
1110 * Returns -1 if we encountered an error preventing us from freeing a
1111 * discarded dcache, or 0 on success.
1114 afs_FreeDiscardedDCache(void)
1117 struct osi_file *tfile;
1120 AFS_STATCNT(afs_FreeDiscardedDCache);
1122 ObtainWriteLock(&afs_xdcache, 510);
1123 if (!afs_blocksDiscarded) {
1124 ReleaseWriteLock(&afs_xdcache);
1129 * Get an entry from the list of discarded cache elements
1131 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1133 ReleaseWriteLock(&afs_xdcache);
1137 afs_discardDCCount--;
1138 size = ((tdc->f.chunkBytes + afs_fsfragsize) ^ afs_fsfragsize) >> 10; /* round up */
1139 afs_blocksDiscarded -= size;
1140 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1141 /* We can lock because we just took it off the free list */
1142 ObtainWriteLock(&tdc->lock, 626);
1143 ReleaseWriteLock(&afs_xdcache);
1146 * Truncate the element to reclaim its space
1148 tfile = afs_CFileOpen(&tdc->f.inode);
1149 afs_CFileTruncate(tfile, 0);
1150 afs_CFileClose(tfile);
1151 afs_AdjustSize(tdc, 0);
1152 afs_DCMoveBucket(tdc, 0, 0);
1155 * Free the element we just truncated
1157 ObtainWriteLock(&afs_xdcache, 511);
1158 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1159 afs_FreeDCache(tdc);
1160 tdc->f.states &= ~(DRO|DBackup|DRW);
1161 ReleaseWriteLock(&tdc->lock);
1163 ReleaseWriteLock(&afs_xdcache);
1169 * Free as many entries from the list of discarded cache elements
1170 * as we need to get the free space down below CM_WAITFORDRAINPCT (98%).
1175 afs_MaybeFreeDiscardedDCache(void)
1178 AFS_STATCNT(afs_MaybeFreeDiscardedDCache);
1180 while (afs_blocksDiscarded
1181 && (afs_blocksUsed >
1182 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
1183 int code = afs_FreeDiscardedDCache();
1185 /* Callers depend on us to get the afs_blocksDiscarded count down.
1186 * If we cannot do that, the callers can spin by calling us over
1187 * and over. Panic for now until we can figure out something
1189 osi_Panic("Error freeing discarded dcache");
1196 * Try to free up a certain number of disk slots.
1198 * \param anumber Targeted number of disk slots to free up.
1200 * \note Environment:
1201 * Must be called with afs_xdcache write-locked.
1205 afs_GetDownDSlot(int anumber)
1207 struct afs_q *tq, *nq;
1212 AFS_STATCNT(afs_GetDownDSlot);
1213 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
1214 osi_Panic("diskless getdowndslot");
1216 if (CheckLock(&afs_xdcache) != -1)
1217 osi_Panic("getdowndslot nolock");
1219 /* decrement anumber first for all dudes in free list */
1220 for (tdc = afs_freeDSList; tdc; tdc = (struct dcache *)tdc->lruq.next)
1223 return; /* enough already free */
1225 for (cnt = 0, tq = afs_DLRU.prev; tq != &afs_DLRU && anumber > 0;
1227 tdc = (struct dcache *)tq; /* q is first elt in dcache entry */
1228 nq = QPrev(tq); /* in case we remove it */
1229 if (tdc->refCount == 0) {
1230 if ((ix = tdc->index) == NULLIDX)
1231 osi_Panic("getdowndslot");
1232 /* pull the entry out of the lruq and put it on the free list */
1233 QRemove(&tdc->lruq);
1235 /* write-through if modified */
1236 if (tdc->dflags & DFEntryMod) {
1237 #if defined(AFS_SGI_ENV) && defined(AFS_SGI_SHORTSTACK)
1239 * ask proxy to do this for us - we don't have the stack space
1241 while (tdc->dflags & DFEntryMod) {
1244 s = SPLOCK(afs_sgibklock);
1245 if (afs_sgibklist == NULL) {
1246 /* if slot is free, grab it. */
1247 afs_sgibklist = tdc;
1248 SV_SIGNAL(&afs_sgibksync);
1250 /* wait for daemon to (start, then) finish. */
1251 SP_WAIT(afs_sgibklock, s, &afs_sgibkwait, PINOD);
1255 tdc->dflags &= ~DFEntryMod;
1256 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
1260 /* finally put the entry in the free list */
1261 afs_indexTable[ix] = NULL;
1262 afs_indexFlags[ix] &= ~IFEverUsed;
1263 tdc->index = NULLIDX;
1264 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
1265 afs_freeDSList = tdc;
1269 } /*afs_GetDownDSlot */
1276 * Increment the reference count on a disk cache entry,
1277 * which already has a non-zero refcount. In order to
1278 * increment the refcount of a zero-reference entry, you
1279 * have to hold afs_xdcache.
1282 * adc : Pointer to the dcache entry to increment.
1285 * Nothing interesting.
1288 afs_RefDCache(struct dcache *adc)
1290 ObtainWriteLock(&adc->tlock, 627);
1291 if (adc->refCount < 0)
1292 osi_Panic("RefDCache: negative refcount");
1294 ReleaseWriteLock(&adc->tlock);
1303 * Decrement the reference count on a disk cache entry.
1306 * ad : Ptr to the dcache entry to decrement.
1309 * Nothing interesting.
1312 afs_PutDCache(struct dcache *adc)
1314 AFS_STATCNT(afs_PutDCache);
1315 ObtainWriteLock(&adc->tlock, 276);
1316 if (adc->refCount <= 0)
1317 osi_Panic("putdcache");
1319 ReleaseWriteLock(&adc->tlock);
1328 * Try to discard all data associated with this file from the
1332 * avc : Pointer to the cache info for the file.
1335 * Both pvnLock and lock are write held.
1338 afs_TryToSmush(struct vcache *avc, afs_ucred_t *acred, int sync)
1343 AFS_STATCNT(afs_TryToSmush);
1344 afs_Trace2(afs_iclSetp, CM_TRACE_TRYTOSMUSH, ICL_TYPE_POINTER, avc,
1345 ICL_TYPE_OFFSET, ICL_HANDLE_OFFSET(avc->f.m.Length));
1346 sync = 1; /* XX Temp testing XX */
1348 #if defined(AFS_SUN5_ENV)
1349 ObtainWriteLock(&avc->vlock, 573);
1350 avc->activeV++; /* block new getpages */
1351 ReleaseWriteLock(&avc->vlock);
1354 /* Flush VM pages */
1355 osi_VM_TryToSmush(avc, acred, sync);
1358 * Get the hash chain containing all dce's for this fid
1360 i = DVHash(&avc->f.fid);
1361 ObtainWriteLock(&afs_xdcache, 277);
1362 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1363 i = afs_dvnextTbl[index]; /* next pointer this hash table */
1364 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1365 int releaseTlock = 1;
1366 tdc = afs_GetValidDSlot(index);
1368 /* afs_TryToSmush is best-effort; we may not actually discard
1369 * everything, so failure to discard a dcache due to an i/o
1373 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1375 if ((afs_indexFlags[index] & IFDataMod) == 0
1376 && tdc->refCount == 1) {
1377 ReleaseReadLock(&tdc->tlock);
1379 afs_FlushDCache(tdc);
1382 afs_indexTable[index] = 0;
1385 ReleaseReadLock(&tdc->tlock);
1389 #if defined(AFS_SUN5_ENV)
1390 ObtainWriteLock(&avc->vlock, 545);
1391 if (--avc->activeV == 0 && (avc->vstates & VRevokeWait)) {
1392 avc->vstates &= ~VRevokeWait;
1393 afs_osi_Wakeup((char *)&avc->vstates);
1395 ReleaseWriteLock(&avc->vlock);
1397 ReleaseWriteLock(&afs_xdcache);
1399 * It's treated like a callback so that when we do lookups we'll
1400 * invalidate the unique bit if any
1401 * trytoSmush occured during the lookup call
1407 * afs_DCacheMissingChunks
1410 * Given the cached info for a file, return the number of chunks that
1411 * are not available from the dcache.
1414 * avc: Pointer to the (held) vcache entry to look in.
1417 * The number of chunks which are not currently cached.
1420 * The vcache entry is held upon entry.
1424 afs_DCacheMissingChunks(struct vcache *avc)
1427 afs_size_t totalLength = 0;
1428 afs_uint32 totalChunks = 0;
1431 totalLength = avc->f.m.Length;
1432 if (avc->f.truncPos < totalLength)
1433 totalLength = avc->f.truncPos;
1435 /* Length is 0, no chunk missing. */
1436 if (totalLength == 0)
1439 /* If totalLength is a multiple of chunksize, the last byte appears
1440 * as being part of the next chunk, which does not exist.
1441 * Decrementing totalLength by one fixes that.
1444 totalChunks = (AFS_CHUNK(totalLength) + 1);
1446 /* If we're a directory, we only ever have one chunk, regardless of
1447 * the size of the dir.
1449 if (avc->f.fid.Fid.Vnode & 1 || vType(avc) == VDIR)
1453 printf("Should have %d chunks for %u bytes\n",
1454 totalChunks, (totalLength + 1));
1456 i = DVHash(&avc->f.fid);
1457 ObtainWriteLock(&afs_xdcache, 1001);
1458 for (index = afs_dvhashTbl[i]; index != NULLIDX; index = i) {
1459 i = afs_dvnextTbl[index];
1460 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1461 tdc = afs_GetValidDSlot(index);
1463 if (!FidCmp(&tdc->f.fid, &avc->f.fid)) {
1466 ReleaseReadLock(&tdc->tlock);
1471 ReleaseWriteLock(&afs_xdcache);
1473 /*printf("Missing %d chunks\n", totalChunks);*/
1475 return (totalChunks);
1482 * Given the cached info for a file and a byte offset into the
1483 * file, make sure the dcache entry for that file and containing
1484 * the given byte is available, returning it to our caller.
1487 * avc : Pointer to the (held) vcache entry to look in.
1488 * abyte : Which byte we want to get to.
1491 * Pointer to the dcache entry covering the file & desired byte,
1492 * or NULL if not found.
1495 * The vcache entry is held upon entry.
1499 afs_FindDCache(struct vcache *avc, afs_size_t abyte)
1503 struct dcache *tdc = NULL;
1505 AFS_STATCNT(afs_FindDCache);
1506 chunk = AFS_CHUNK(abyte);
1509 * Hash on the [fid, chunk] and get the corresponding dcache index
1510 * after write-locking the dcache.
1512 i = DCHash(&avc->f.fid, chunk);
1513 ObtainWriteLock(&afs_xdcache, 278);
1514 for (index = afs_dchashTbl[i]; index != NULLIDX; index = afs_dcnextTbl[index]) {
1515 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1516 tdc = afs_GetValidDSlot(index);
1518 /* afs_FindDCache is best-effort; we may not find the given
1519 * file/offset, so if we cannot find the given dcache due to
1520 * i/o errors, that is okay. */
1523 ReleaseReadLock(&tdc->tlock);
1524 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1525 break; /* leaving refCount high for caller */
1530 if (index != NULLIDX) {
1531 hset(afs_indexTimes[tdc->index], afs_indexCounter);
1532 hadd32(afs_indexCounter, 1);
1533 ReleaseWriteLock(&afs_xdcache);
1536 ReleaseWriteLock(&afs_xdcache);
1538 } /*afs_FindDCache */
1540 /* only call these from afs_AllocDCache() */
1541 static struct dcache *
1542 afs_AllocFreeDSlot(void)
1546 tdc = afs_GetDSlotFromList(&afs_freeDCList);
1550 afs_indexFlags[tdc->index] &= ~IFFree;
1551 ObtainWriteLock(&tdc->lock, 604);
1556 static struct dcache *
1557 afs_AllocDiscardDSlot(afs_int32 lock)
1560 afs_uint32 size = 0;
1561 struct osi_file *file;
1563 tdc = afs_GetDSlotFromList(&afs_discardDCList);
1567 afs_indexFlags[tdc->index] &= ~IFDiscarded;
1568 ObtainWriteLock(&tdc->lock, 605);
1569 afs_discardDCCount--;
1571 ((tdc->f.chunkBytes +
1572 afs_fsfragsize) ^ afs_fsfragsize) >> 10;
1573 tdc->f.states &= ~(DRO|DBackup|DRW);
1574 afs_DCMoveBucket(tdc, size, 0);
1575 afs_blocksDiscarded -= size;
1576 afs_stats_cmperf.cacheBlocksDiscarded = afs_blocksDiscarded;
1578 /* Truncate the chunk so zeroes get filled properly */
1579 file = afs_CFileOpen(&tdc->f.inode);
1580 afs_CFileTruncate(file, 0);
1581 afs_CFileClose(file);
1582 afs_AdjustSize(tdc, 0);
1589 * Get a fresh dcache from the free or discarded list.
1591 * \param avc Who's dcache is this going to be?
1592 * \param chunk The position where it will be placed in.
1593 * \param lock How are locks held.
1594 * \param ashFid If this dcache going to be used for a shadow dir,
1597 * \note Required locks:
1599 * - avc (R if (lock & 1) set and W otherwise)
1600 * \note It write locks the new dcache. The caller must unlock it.
1602 * \return The new dcache.
1605 afs_AllocDCache(struct vcache *avc, afs_int32 chunk, afs_int32 lock,
1606 struct VenusFid *ashFid)
1608 struct dcache *tdc = NULL;
1610 /* if (lock & 2), prefer 'free' dcaches; otherwise, prefer 'discard'
1611 * dcaches. In either case, try both if our first choice doesn't work. */
1613 tdc = afs_AllocFreeDSlot();
1615 tdc = afs_AllocDiscardDSlot(lock);
1618 tdc = afs_AllocDiscardDSlot(lock);
1620 tdc = afs_AllocFreeDSlot();
1629 * avc->lock(R) if setLocks
1630 * avc->lock(W) if !setLocks
1636 * Fill in the newly-allocated dcache record.
1638 afs_indexFlags[tdc->index] &= ~(IFDirtyPages | IFAnyPages);
1640 /* Use shadow fid if provided. */
1641 tdc->f.fid = *ashFid;
1643 /* Use normal vcache's fid otherwise. */
1644 tdc->f.fid = avc->f.fid;
1645 if (avc->f.states & CRO)
1646 tdc->f.states = DRO;
1647 else if (avc->f.states & CBackup)
1648 tdc->f.states = DBackup;
1650 tdc->f.states = DRW;
1651 afs_DCMoveBucket(tdc, 0, afs_DCGetBucket(avc));
1652 afs_indexUnique[tdc->index] = tdc->f.fid.Fid.Unique;
1654 hones(tdc->f.versionNo); /* invalid value */
1655 tdc->f.chunk = chunk;
1656 tdc->validPos = AFS_CHUNKTOBASE(chunk);
1658 if (tdc->lruq.prev == &tdc->lruq)
1659 osi_Panic("lruq 1");
1668 * This function is called to obtain a reference to data stored in
1669 * the disk cache, locating a chunk of data containing the desired
1670 * byte and returning a reference to the disk cache entry, with its
1671 * reference count incremented.
1675 * avc : Ptr to a vcache entry (unlocked)
1676 * abyte : Byte position in the file desired
1677 * areq : Request structure identifying the requesting user.
1678 * aflags : Settings as follows:
1680 * 2 : Return after creating entry.
1681 * 4 : called from afs_vnop_write.c
1682 * *alen contains length of data to be written.
1684 * aoffset : Set to the offset within the chunk where the resident
1686 * alen : Set to the number of bytes of data after the desired
1687 * byte (including the byte itself) which can be read
1691 * The vcache entry pointed to by avc is unlocked upon entry.
1695 * Update the vnode-to-dcache hint if we can get the vnode lock
1696 * right away. Assumes dcache entry is at least read-locked.
1699 updateV2DC(int lockVc, struct vcache *v, struct dcache *d, int src)
1701 if (!lockVc || 0 == NBObtainWriteLock(&v->lock, src)) {
1702 if (hsame(v->f.m.DataVersion, d->f.versionNo) && v->callback)
1705 ReleaseWriteLock(&v->lock);
1709 /* avc - Write-locked unless aflags & 1 */
1711 afs_GetDCache(struct vcache *avc, afs_size_t abyte,
1712 struct vrequest *areq, afs_size_t * aoffset,
1713 afs_size_t * alen, int aflags)
1715 afs_int32 i, code, shortcut;
1716 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
1717 afs_int32 adjustsize = 0;
1723 afs_size_t Position = 0;
1724 afs_int32 size, tlen; /* size of segment to transfer */
1725 struct afs_FetchOutput *tsmall = 0;
1727 struct osi_file *file;
1728 struct afs_conn *tc;
1730 struct server *newCallback = NULL;
1731 char setNewCallback;
1732 char setVcacheStatus;
1733 char doVcacheUpdate;
1735 int doAdjustSize = 0;
1736 int doReallyAdjustSize = 0;
1737 int overWriteWholeChunk = 0;
1738 struct rx_connection *rxconn;
1741 struct afs_stats_AccessInfo *accP; /*Ptr to access record in stats */
1742 int fromReplica; /*Are we reading from a replica? */
1743 int numFetchLoops; /*# times around the fetch/analyze loop */
1744 #endif /* AFS_NOSTATS */
1746 AFS_STATCNT(afs_GetDCache);
1750 setLocks = aflags & 1;
1753 * Determine the chunk number and offset within the chunk corresponding
1754 * to the desired byte.
1756 if (avc->f.fid.Fid.Vnode & 1) { /* if (vType(avc) == VDIR) */
1759 chunk = AFS_CHUNK(abyte);
1762 /* come back to here if we waited for the cache to drain. */
1765 setNewCallback = setVcacheStatus = 0;
1769 ObtainWriteLock(&avc->lock, 616);
1771 ObtainReadLock(&avc->lock);
1776 * avc->lock(R) if setLocks && !slowPass
1777 * avc->lock(W) if !setLocks || slowPass
1782 /* check hints first! (might could use bcmp or some such...) */
1783 if ((tdc = avc->dchint)) {
1787 * The locking order between afs_xdcache and dcache lock matters.
1788 * The hint dcache entry could be anywhere, even on the free list.
1789 * Locking afs_xdcache ensures that noone is trying to pull dcache
1790 * entries from the free list, and thereby assuming them to be not
1791 * referenced and not locked.
1793 ObtainReadLock(&afs_xdcache);
1794 dcLocked = (0 == NBObtainSharedLock(&tdc->lock, 601));
1796 if (dcLocked && (tdc->index != NULLIDX)
1797 && !FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk
1798 && !(afs_indexFlags[tdc->index] & (IFFree | IFDiscarded))) {
1799 /* got the right one. It might not be the right version, and it
1800 * might be fetching, but it's the right dcache entry.
1802 /* All this code should be integrated better with what follows:
1803 * I can save a good bit more time under a write lock if I do..
1805 ObtainWriteLock(&tdc->tlock, 603);
1807 ReleaseWriteLock(&tdc->tlock);
1809 ReleaseReadLock(&afs_xdcache);
1812 if (hsame(tdc->f.versionNo, avc->f.m.DataVersion)
1813 && !(tdc->dflags & DFFetching)) {
1815 afs_stats_cmperf.dcacheHits++;
1816 ObtainWriteLock(&afs_xdcache, 559);
1817 QRemove(&tdc->lruq);
1818 QAdd(&afs_DLRU, &tdc->lruq);
1819 ReleaseWriteLock(&afs_xdcache);
1822 * avc->lock(R) if setLocks && !slowPass
1823 * avc->lock(W) if !setLocks || slowPass
1830 ReleaseSharedLock(&tdc->lock);
1831 ReleaseReadLock(&afs_xdcache);
1839 * avc->lock(R) if setLocks && !slowPass
1840 * avc->lock(W) if !setLocks || slowPass
1841 * tdc->lock(S) if tdc
1844 if (!tdc) { /* If the hint wasn't the right dcache entry */
1845 int dslot_error = 0;
1847 * Hash on the [fid, chunk] and get the corresponding dcache index
1848 * after write-locking the dcache.
1853 * avc->lock(R) if setLocks && !slowPass
1854 * avc->lock(W) if !setLocks || slowPass
1857 i = DCHash(&avc->f.fid, chunk);
1858 /* check to make sure our space is fine */
1859 afs_MaybeWakeupTruncateDaemon();
1861 ObtainWriteLock(&afs_xdcache, 280);
1863 for (index = afs_dchashTbl[i]; index != NULLIDX; us = index, index = afs_dcnextTbl[index]) {
1864 if (afs_indexUnique[index] == avc->f.fid.Fid.Unique) {
1865 tdc = afs_GetValidDSlot(index);
1867 /* we got an i/o error when trying to get the given dslot,
1868 * but do not bail out just yet; it is possible the dcache
1869 * we're looking for is elsewhere, so it doesn't matter if
1870 * we can't load this one. */
1874 ReleaseReadLock(&tdc->tlock);
1877 * avc->lock(R) if setLocks && !slowPass
1878 * avc->lock(W) if !setLocks || slowPass
1881 if (!FidCmp(&tdc->f.fid, &avc->f.fid) && chunk == tdc->f.chunk) {
1882 /* Move it up in the beginning of the list */
1883 if (afs_dchashTbl[i] != index) {
1884 afs_dcnextTbl[us] = afs_dcnextTbl[index];
1885 afs_dcnextTbl[index] = afs_dchashTbl[i];
1886 afs_dchashTbl[i] = index;
1888 ReleaseWriteLock(&afs_xdcache);
1889 ObtainSharedLock(&tdc->lock, 606);
1890 break; /* leaving refCount high for caller */
1898 * If we didn't find the entry, we'll create one.
1900 if (index == NULLIDX) {
1903 * avc->lock(R) if setLocks
1904 * avc->lock(W) if !setLocks
1907 afs_Trace2(afs_iclSetp, CM_TRACE_GETDCACHE1, ICL_TYPE_POINTER,
1908 avc, ICL_TYPE_INT32, chunk);
1911 /* We couldn't find the dcache we want, but we hit some i/o
1912 * errors when trying to find it, so we're not sure if the
1913 * dcache we want is in the cache or not. Error out, so we
1914 * don't try to possibly create 2 separate dcaches for the
1915 * same exact data. */
1916 ReleaseWriteLock(&afs_xdcache);
1920 if (afs_discardDCList == NULLIDX && afs_freeDCList == NULLIDX) {
1922 avc->f.states |= CDCLock;
1923 /* just need slots */
1924 afs_GetDownD(5, (int *)0, afs_DCGetBucket(avc));
1926 avc->f.states &= ~CDCLock;
1928 tdc = afs_AllocDCache(avc, chunk, aflags, NULL);
1930 /* If we can't get space for 5 mins we give up and panic */
1931 if (++downDCount > 300)
1932 osi_Panic("getdcache");
1933 ReleaseWriteLock(&afs_xdcache);
1936 * avc->lock(R) if setLocks
1937 * avc->lock(W) if !setLocks
1939 afs_osi_Wait(1000, 0, 0);
1945 * avc->lock(R) if setLocks
1946 * avc->lock(W) if !setLocks
1952 * Now add to the two hash chains - note that i is still set
1953 * from the above DCHash call.
1955 afs_dcnextTbl[tdc->index] = afs_dchashTbl[i];
1956 afs_dchashTbl[i] = tdc->index;
1957 i = DVHash(&avc->f.fid);
1958 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[i];
1959 afs_dvhashTbl[i] = tdc->index;
1960 tdc->dflags = DFEntryMod;
1962 afs_MaybeWakeupTruncateDaemon();
1963 ReleaseWriteLock(&afs_xdcache);
1964 ConvertWToSLock(&tdc->lock);
1969 /* vcache->dcache hint failed */
1972 * avc->lock(R) if setLocks && !slowPass
1973 * avc->lock(W) if !setLocks || slowPass
1976 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE2, ICL_TYPE_POINTER, avc,
1977 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
1978 hgetlo(tdc->f.versionNo), ICL_TYPE_INT32,
1979 hgetlo(avc->f.m.DataVersion));
1981 * Here we have the entry in tdc, with its refCount incremented.
1982 * Note: we don't use the S-lock on avc; it costs concurrency when
1983 * storing a file back to the server.
1987 * Not a newly created file so we need to check the file's length and
1988 * compare data versions since someone could have changed the data or we're
1989 * reading a file written elsewhere. We only want to bypass doing no-op
1990 * read rpcs on newly created files (dv of 0) since only then we guarantee
1991 * that this chunk's data hasn't been filled by another client.
1993 size = AFS_CHUNKSIZE(abyte);
1994 if (aflags & 4) /* called from write */
1996 else /* called from read */
1997 tlen = tdc->validPos - abyte;
1998 Position = AFS_CHUNKTOBASE(chunk);
1999 afs_Trace4(afs_iclSetp, CM_TRACE_GETDCACHE3, ICL_TYPE_INT32, tlen,
2000 ICL_TYPE_INT32, aflags, ICL_TYPE_OFFSET,
2001 ICL_HANDLE_OFFSET(abyte), ICL_TYPE_OFFSET,
2002 ICL_HANDLE_OFFSET(Position));
2003 if ((aflags & 4) && (hiszero(avc->f.m.DataVersion)))
2005 if ((AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length) ||
2006 ((aflags & 4) && (abyte == Position) && (tlen >= size)))
2007 overWriteWholeChunk = 1;
2008 if (doAdjustSize || overWriteWholeChunk) {
2009 #if defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV)
2011 #ifdef AFS_SGI64_ENV
2014 #else /* AFS_SGI64_ENV */
2017 #endif /* AFS_SGI64_ENV */
2018 #else /* AFS_SGI_ENV */
2021 #endif /* AFS_SGI_ENV */
2022 if (AFS_CHUNKTOBASE(chunk) + adjustsize >= avc->f.m.Length &&
2023 #else /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2024 #if defined(AFS_SUN5_ENV)
2025 if ((doAdjustSize || (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length)) &&
2027 if (AFS_CHUNKTOBASE(chunk) >= avc->f.m.Length &&
2029 #endif /* defined(AFS_AIX32_ENV) || defined(AFS_SGI_ENV) */
2030 !hsame(avc->f.m.DataVersion, tdc->f.versionNo))
2031 doReallyAdjustSize = 1;
2033 if (doReallyAdjustSize || overWriteWholeChunk) {
2034 /* no data in file to read at this position */
2035 UpgradeSToWLock(&tdc->lock, 607);
2036 file = afs_CFileOpen(&tdc->f.inode);
2037 afs_CFileTruncate(file, 0);
2038 afs_CFileClose(file);
2039 afs_AdjustSize(tdc, 0);
2040 hset(tdc->f.versionNo, avc->f.m.DataVersion);
2041 tdc->dflags |= DFEntryMod;
2043 ConvertWToSLock(&tdc->lock);
2048 * We must read in the whole chunk if the version number doesn't
2052 /* don't need data, just a unique dcache entry */
2053 ObtainWriteLock(&afs_xdcache, 608);
2054 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2055 hadd32(afs_indexCounter, 1);
2056 ReleaseWriteLock(&afs_xdcache);
2058 updateV2DC(setLocks, avc, tdc, 553);
2059 if (vType(avc) == VDIR)
2062 *aoffset = AFS_CHUNKOFFSET(abyte);
2063 if (tdc->validPos < abyte)
2064 *alen = (afs_size_t) 0;
2066 *alen = tdc->validPos - abyte;
2067 ReleaseSharedLock(&tdc->lock);
2070 ReleaseWriteLock(&avc->lock);
2072 ReleaseReadLock(&avc->lock);
2074 return tdc; /* check if we're done */
2079 * avc->lock(R) if setLocks && !slowPass
2080 * avc->lock(W) if !setLocks || slowPass
2083 osi_Assert((setLocks && !slowPass) || WriteLocked(&avc->lock));
2085 setNewCallback = setVcacheStatus = 0;
2089 * avc->lock(R) if setLocks && !slowPass
2090 * avc->lock(W) if !setLocks || slowPass
2093 if (!hsame(avc->f.m.DataVersion, tdc->f.versionNo) && !overWriteWholeChunk) {
2095 * Version number mismatch.
2098 * If we are disconnected, then we can't do much of anything
2099 * because the data doesn't match the file.
2101 if (AFS_IS_DISCONNECTED) {
2102 ReleaseSharedLock(&tdc->lock);
2105 ReleaseWriteLock(&avc->lock);
2107 ReleaseReadLock(&avc->lock);
2109 /* Flush the Dcache */
2114 UpgradeSToWLock(&tdc->lock, 609);
2117 * If data ever existed for this vnode, and this is a text object,
2118 * do some clearing. Now, you'd think you need only do the flush
2119 * when VTEXT is on, but VTEXT is turned off when the text object
2120 * is freed, while pages are left lying around in memory marked
2121 * with this vnode. If we would reactivate (create a new text
2122 * object from) this vnode, we could easily stumble upon some of
2123 * these old pages in pagein. So, we always flush these guys.
2124 * Sun has a wonderful lack of useful invariants in this system.
2126 * avc->flushDV is the data version # of the file at the last text
2127 * flush. Clearly, at least, we don't have to flush the file more
2128 * often than it changes
2130 if (hcmp(avc->flushDV, avc->f.m.DataVersion) < 0) {
2132 * By here, the cache entry is always write-locked. We can
2133 * deadlock if we call osi_Flush with the cache entry locked...
2134 * Unlock the dcache too.
2136 ReleaseWriteLock(&tdc->lock);
2137 if (setLocks && !slowPass)
2138 ReleaseReadLock(&avc->lock);
2140 ReleaseWriteLock(&avc->lock);
2144 * Call osi_FlushPages in open, read/write, and map, since it
2145 * is too hard here to figure out if we should lock the
2148 if (setLocks && !slowPass)
2149 ObtainReadLock(&avc->lock);
2151 ObtainWriteLock(&avc->lock, 66);
2152 ObtainWriteLock(&tdc->lock, 610);
2157 * avc->lock(R) if setLocks && !slowPass
2158 * avc->lock(W) if !setLocks || slowPass
2162 /* Watch for standard race condition around osi_FlushText */
2163 if (hsame(avc->f.m.DataVersion, tdc->f.versionNo)) {
2164 updateV2DC(setLocks, avc, tdc, 569); /* set hint */
2165 afs_stats_cmperf.dcacheHits++;
2166 ConvertWToSLock(&tdc->lock);
2170 /* Sleep here when cache needs to be drained. */
2171 if (setLocks && !slowPass
2172 && (afs_blocksUsed >
2173 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks))) {
2174 /* Make sure truncate daemon is running */
2175 afs_MaybeWakeupTruncateDaemon();
2176 ObtainWriteLock(&tdc->tlock, 614);
2177 tdc->refCount--; /* we'll re-obtain the dcache when we re-try. */
2178 ReleaseWriteLock(&tdc->tlock);
2179 ReleaseWriteLock(&tdc->lock);
2180 ReleaseReadLock(&avc->lock);
2181 while ((afs_blocksUsed - afs_blocksDiscarded) >
2182 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
2183 afs_WaitForCacheDrain = 1;
2184 afs_osi_Sleep(&afs_WaitForCacheDrain);
2186 afs_MaybeFreeDiscardedDCache();
2187 /* need to check if someone else got the chunk first. */
2188 goto RetryGetDCache;
2191 Position = AFS_CHUNKBASE(abyte);
2192 if (vType(avc) == VDIR) {
2193 size = avc->f.m.Length;
2194 if (size > tdc->f.chunkBytes) {
2195 /* pre-reserve space for file */
2196 afs_AdjustSize(tdc, size);
2198 size = 999999999; /* max size for transfer */
2200 afs_size_t maxGoodLength;
2202 /* estimate how much data we're expecting back from the server,
2203 * and reserve space in the dcache entry for it */
2205 maxGoodLength = avc->f.m.Length;
2206 if (avc->f.truncPos < maxGoodLength)
2207 maxGoodLength = avc->f.truncPos;
2209 size = AFS_CHUNKSIZE(abyte); /* expected max size */
2210 if (Position + size > maxGoodLength)
2211 size = maxGoodLength - Position;
2213 size = 0; /* Handle random races */
2214 if (size > tdc->f.chunkBytes) {
2215 /* pre-reserve estimated space for file */
2216 afs_AdjustSize(tdc, size); /* changes chunkBytes */
2220 /* For the actual fetch, do not limit the request to the
2221 * length of the file. If this results in a read past EOF on
2222 * the server, the server will just reply with less data than
2223 * requested. If we limit ourselves to only requesting data up
2224 * to the avc file length, we open ourselves up to races if the
2225 * file is extended on the server at about the same time.
2227 * However, we must restrict ourselves to the avc->f.truncPos
2228 * length, since this represents an outstanding local
2229 * truncation of the file that will be committed to the
2230 * fileserver when we actually write the fileserver contents.
2231 * If we do not restrict the fetch length based on
2232 * avc->f.truncPos, a different truncate operation extending
2233 * the file length could cause the old data after
2234 * avc->f.truncPos to reappear, instead of extending the file
2235 * with NUL bytes. */
2236 size = AFS_CHUNKSIZE(abyte);
2237 if (Position + size > avc->f.truncPos) {
2238 size = avc->f.truncPos - Position;
2245 if (afs_mariner && !tdc->f.chunk)
2246 afs_MarinerLog("fetch$Fetching", avc); /* , Position, size, afs_indexCounter ); */
2248 * Right now, we only have one tool, and it's a hammer. So, we
2249 * fetch the whole file.
2251 DZap(tdc); /* pages in cache may be old */
2252 file = afs_CFileOpen(&tdc->f.inode);
2253 afs_RemoveVCB(&avc->f.fid);
2254 tdc->f.states |= DWriting;
2255 tdc->dflags |= DFFetching;
2256 tdc->validPos = Position; /* which is AFS_CHUNKBASE(abyte) */
2257 if (tdc->mflags & DFFetchReq) {
2258 tdc->mflags &= ~DFFetchReq;
2259 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2260 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2261 __FILE__, ICL_TYPE_INT32, __LINE__,
2262 ICL_TYPE_POINTER, tdc, ICL_TYPE_INT32,
2265 tsmall = osi_AllocLargeSpace(sizeof(struct afs_FetchOutput));
2266 setVcacheStatus = 0;
2269 * Remember if we are doing the reading from a replicated volume,
2270 * and how many times we've zipped around the fetch/analyze loop.
2272 fromReplica = (avc->f.states & CRO) ? 1 : 0;
2274 accP = &(afs_stats_cmfullperf.accessinf);
2276 (accP->replicatedRefs)++;
2278 (accP->unreplicatedRefs)++;
2279 #endif /* AFS_NOSTATS */
2280 /* this is a cache miss */
2281 afs_Trace4(afs_iclSetp, CM_TRACE_FETCHPROC, ICL_TYPE_POINTER, avc,
2282 ICL_TYPE_FID, &(avc->f.fid), ICL_TYPE_OFFSET,
2283 ICL_HANDLE_OFFSET(Position), ICL_TYPE_INT32, size);
2286 afs_stats_cmperf.dcacheMisses++;
2289 * Dynamic root support: fetch data from local memory.
2291 if (afs_IsDynroot(avc)) {
2295 afs_GetDynroot(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2297 dynrootDir += Position;
2298 dynrootLen -= Position;
2299 if (size > dynrootLen)
2303 code = afs_CFileWrite(file, 0, dynrootDir, size);
2311 tdc->validPos = Position + size;
2312 afs_CFileTruncate(file, size); /* prune it */
2313 } else if (afs_IsDynrootMount(avc)) {
2317 afs_GetDynrootMount(&dynrootDir, &dynrootLen, &tsmall->OutStatus);
2319 dynrootDir += Position;
2320 dynrootLen -= Position;
2321 if (size > dynrootLen)
2325 code = afs_CFileWrite(file, 0, dynrootDir, size);
2333 tdc->validPos = Position + size;
2334 afs_CFileTruncate(file, size); /* prune it */
2337 * Not a dynamic vnode: do the real fetch.
2342 * avc->lock(R) if setLocks && !slowPass
2343 * avc->lock(W) if !setLocks || slowPass
2347 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK, &rxconn);
2352 (accP->numReplicasAccessed)++;
2354 #endif /* AFS_NOSTATS */
2355 if (!setLocks || slowPass) {
2356 avc->callback = tc->parent->srvr->server;
2358 newCallback = tc->parent->srvr->server;
2362 code = afs_CacheFetchProc(tc, rxconn, file, Position, tdc,
2368 /* callback could have been broken (or expired) in a race here,
2369 * but we return the data anyway. It's as good as we knew about
2370 * when we started. */
2372 * validPos is updated by CacheFetchProc, and can only be
2373 * modifed under a dcache write lock, which we've blocked out
2375 size = tdc->validPos - Position; /* actual segment size */
2378 afs_CFileTruncate(file, size); /* prune it */
2380 if (!setLocks || slowPass) {
2381 ObtainWriteLock(&afs_xcbhash, 453);
2382 afs_DequeueCallback(avc);
2383 avc->f.states &= ~(CStatd | CUnique);
2384 avc->callback = NULL;
2385 ReleaseWriteLock(&afs_xcbhash);
2386 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2387 osi_dnlc_purgedp(avc);
2389 /* Something lost. Forget about performance, and go
2390 * back with a vcache write lock.
2392 afs_CFileTruncate(file, 0);
2393 afs_AdjustSize(tdc, 0);
2394 afs_CFileClose(file);
2395 osi_FreeLargeSpace(tsmall);
2397 ReleaseWriteLock(&tdc->lock);
2400 ReleaseReadLock(&avc->lock);
2402 goto RetryGetDCache;
2406 } while (afs_Analyze
2407 (tc, rxconn, code, &avc->f.fid, areq,
2408 AFS_STATS_FS_RPCIDX_FETCHDATA, SHARED_LOCK, NULL));
2412 * avc->lock(R) if setLocks && !slowPass
2413 * avc->lock(W) if !setLocks || slowPass
2419 * In the case of replicated access, jot down info on the number of
2420 * attempts it took before we got through or gave up.
2423 if (numFetchLoops <= 1)
2424 (accP->refFirstReplicaOK)++;
2425 if (numFetchLoops > accP->maxReplicasPerRef)
2426 accP->maxReplicasPerRef = numFetchLoops;
2428 #endif /* AFS_NOSTATS */
2430 tdc->dflags &= ~DFFetching;
2431 if (afs_osi_Wakeup(&tdc->validPos) == 0)
2432 afs_Trace4(afs_iclSetp, CM_TRACE_DCACHEWAKE, ICL_TYPE_STRING,
2433 __FILE__, ICL_TYPE_INT32, __LINE__, ICL_TYPE_POINTER,
2434 tdc, ICL_TYPE_INT32, tdc->dflags);
2435 if (avc->execsOrWriters == 0)
2436 tdc->f.states &= ~DWriting;
2438 /* now, if code != 0, we have an error and should punt.
2439 * note that we have the vcache write lock, either because
2440 * !setLocks or slowPass.
2443 afs_CFileTruncate(file, 0);
2444 afs_AdjustSize(tdc, 0);
2445 afs_CFileClose(file);
2446 ZapDCE(tdc); /* sets DFEntryMod */
2447 if (vType(avc) == VDIR) {
2450 tdc->f.states &= ~(DRO|DBackup|DRW);
2451 afs_DCMoveBucket(tdc, 0, 0);
2452 ReleaseWriteLock(&tdc->lock);
2454 if (!afs_IsDynroot(avc)) {
2455 ObtainWriteLock(&afs_xcbhash, 454);
2456 afs_DequeueCallback(avc);
2457 avc->f.states &= ~(CStatd | CUnique);
2458 ReleaseWriteLock(&afs_xcbhash);
2459 if (avc->f.fid.Fid.Vnode & 1 || (vType(avc) == VDIR))
2460 osi_dnlc_purgedp(avc);
2463 * avc->lock(W); assert(!setLocks || slowPass)
2465 osi_Assert(!setLocks || slowPass);
2471 /* otherwise we copy in the just-fetched info */
2472 afs_CFileClose(file);
2473 afs_AdjustSize(tdc, size); /* new size */
2475 * Copy appropriate fields into vcache. Status is
2476 * copied later where we selectively acquire the
2477 * vcache write lock.
2480 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2482 setVcacheStatus = 1;
2483 hset64(tdc->f.versionNo, tsmall->OutStatus.dataVersionHigh,
2484 tsmall->OutStatus.DataVersion);
2485 tdc->dflags |= DFEntryMod;
2486 afs_indexFlags[tdc->index] |= IFEverUsed;
2487 ConvertWToSLock(&tdc->lock);
2488 } /*Data version numbers don't match */
2491 * Data version numbers match.
2493 afs_stats_cmperf.dcacheHits++;
2494 } /*Data version numbers match */
2496 updateV2DC(setLocks, avc, tdc, 335); /* set hint */
2500 * avc->lock(R) if setLocks && !slowPass
2501 * avc->lock(W) if !setLocks || slowPass
2502 * tdc->lock(S) if tdc
2506 * See if this was a reference to a file in the local cell.
2508 if (afs_IsPrimaryCellNum(avc->f.fid.Cell))
2509 afs_stats_cmperf.dlocalAccesses++;
2511 afs_stats_cmperf.dremoteAccesses++;
2513 /* Fix up LRU info */
2516 ObtainWriteLock(&afs_xdcache, 602);
2517 hset(afs_indexTimes[tdc->index], afs_indexCounter);
2518 hadd32(afs_indexCounter, 1);
2519 ReleaseWriteLock(&afs_xdcache);
2521 /* return the data */
2522 if (vType(avc) == VDIR)
2525 *aoffset = AFS_CHUNKOFFSET(abyte);
2526 *alen = (tdc->f.chunkBytes - *aoffset);
2527 ReleaseSharedLock(&tdc->lock);
2532 * avc->lock(R) if setLocks && !slowPass
2533 * avc->lock(W) if !setLocks || slowPass
2536 /* Fix up the callback and status values in the vcache */
2538 if (setLocks && !slowPass) {
2541 * This is our dirty little secret to parallel fetches.
2542 * We don't write-lock the vcache while doing the fetch,
2543 * but potentially we'll need to update the vcache after
2544 * the fetch is done.
2546 * Drop the read lock and try to re-obtain the write
2547 * lock. If the vcache still has the same DV, it's
2548 * ok to go ahead and install the new data.
2550 afs_hyper_t currentDV, statusDV;
2552 hset(currentDV, avc->f.m.DataVersion);
2554 if (setNewCallback && avc->callback != newCallback)
2558 hset64(statusDV, tsmall->OutStatus.dataVersionHigh,
2559 tsmall->OutStatus.DataVersion);
2561 if (setVcacheStatus && avc->f.m.Length != tsmall->OutStatus.Length)
2563 if (setVcacheStatus && !hsame(currentDV, statusDV))
2567 ReleaseReadLock(&avc->lock);
2569 if (doVcacheUpdate) {
2570 ObtainWriteLock(&avc->lock, 615);
2571 if (!hsame(avc->f.m.DataVersion, currentDV)) {
2572 /* We lose. Someone will beat us to it. */
2574 ReleaseWriteLock(&avc->lock);
2579 /* With slow pass, we've already done all the updates */
2581 ReleaseWriteLock(&avc->lock);
2584 /* Check if we need to perform any last-minute fixes with a write-lock */
2585 if (!setLocks || doVcacheUpdate) {
2587 avc->callback = newCallback;
2588 if (tsmall && setVcacheStatus)
2589 afs_ProcessFS(avc, &tsmall->OutStatus, areq);
2591 ReleaseWriteLock(&avc->lock);
2595 osi_FreeLargeSpace(tsmall);
2598 } /*afs_GetDCache */
2602 * afs_WriteThroughDSlots
2605 * Sweep through the dcache slots and write out any modified
2606 * in-memory data back on to our caching store.
2612 * The afs_xdcache is write-locked through this whole affair.
2615 afs_WriteThroughDSlots(void)
2618 afs_int32 i, touchedit = 0;
2620 struct afs_q DirtyQ, *tq;
2622 AFS_STATCNT(afs_WriteThroughDSlots);
2625 * Because of lock ordering, we can't grab dcache locks while
2626 * holding afs_xdcache. So we enter xdcache, get a reference
2627 * for every dcache entry, and exit xdcache.
2629 ObtainWriteLock(&afs_xdcache, 283);
2631 for (i = 0; i < afs_cacheFiles; i++) {
2632 tdc = afs_indexTable[i];
2634 /* Grab tlock in case the existing refcount isn't zero */
2635 if (tdc && !(afs_indexFlags[i] & (IFFree | IFDiscarded))) {
2636 ObtainWriteLock(&tdc->tlock, 623);
2638 ReleaseWriteLock(&tdc->tlock);
2640 QAdd(&DirtyQ, &tdc->dirty);
2643 ReleaseWriteLock(&afs_xdcache);
2646 * Now, for each dcache entry we found, check if it's dirty.
2647 * If so, get write-lock, get afs_xdcache, which protects
2648 * afs_cacheInodep, and flush it. Don't forget to put back
2652 #define DQTODC(q) ((struct dcache *)(((char *) (q)) - sizeof(struct afs_q)))
2654 for (tq = DirtyQ.prev; tq != &DirtyQ; tq = QPrev(tq)) {
2656 if (tdc->dflags & DFEntryMod) {
2659 wrLock = (0 == NBObtainWriteLock(&tdc->lock, 619));
2661 /* Now that we have the write lock, double-check */
2662 if (wrLock && (tdc->dflags & DFEntryMod)) {
2663 tdc->dflags &= ~DFEntryMod;
2664 ObtainWriteLock(&afs_xdcache, 620);
2665 osi_Assert(afs_WriteDCache(tdc, 1) == 0);
2666 ReleaseWriteLock(&afs_xdcache);
2670 ReleaseWriteLock(&tdc->lock);
2676 ObtainWriteLock(&afs_xdcache, 617);
2677 if (!touchedit && (cacheDiskType != AFS_FCACHE_TYPE_MEM)) {
2678 /* Touch the file to make sure that the mtime on the file is kept
2679 * up-to-date to avoid losing cached files on cold starts because
2680 * their mtime seems old...
2682 struct afs_fheader theader;
2684 theader.magic = AFS_FHMAGIC;
2685 theader.firstCSize = AFS_FIRSTCSIZE;
2686 theader.otherCSize = AFS_OTHERCSIZE;
2687 theader.version = AFS_CI_VERSION;
2688 theader.dataSize = sizeof(struct fcache);
2689 afs_osi_Write(afs_cacheInodep, 0, &theader, sizeof(theader));
2691 ReleaseWriteLock(&afs_xdcache);
2698 * Return a pointer to an freshly initialized dcache entry using
2699 * a memory-based cache. The tlock will be read-locked.
2702 * aslot : Dcache slot to look at.
2703 * needvalid : Whether the specified slot should already exist
2706 * Must be called with afs_xdcache write-locked.
2710 afs_MemGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2715 AFS_STATCNT(afs_MemGetDSlot);
2716 if (CheckLock(&afs_xdcache) != -1)
2717 osi_Panic("getdslot nolock");
2718 if (aslot < 0 || aslot >= afs_cacheFiles)
2719 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2720 tdc = afs_indexTable[aslot];
2722 QRemove(&tdc->lruq); /* move to queue head */
2723 QAdd(&afs_DLRU, &tdc->lruq);
2724 /* We're holding afs_xdcache, but get tlock in case refCount != 0 */
2725 ObtainWriteLock(&tdc->tlock, 624);
2727 ConvertWToRLock(&tdc->tlock);
2731 /* if 'indexvalid' is true, the slot must already exist and be populated
2732 * somewhere. for memcache, the only place that dcache entries exist is
2733 * in memory, so if we did not find it above, something is very wrong. */
2734 osi_Assert(!indexvalid);
2736 if (!afs_freeDSList)
2737 afs_GetDownDSlot(4);
2738 if (!afs_freeDSList) {
2739 /* none free, making one is better than a panic */
2740 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2741 tdc = afs_osi_Alloc(sizeof(struct dcache));
2742 osi_Assert(tdc != NULL);
2743 #ifdef KERNEL_HAVE_PIN
2744 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2747 tdc = afs_freeDSList;
2748 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2751 tdc->dflags = 0; /* up-to-date, not in free q */
2753 QAdd(&afs_DLRU, &tdc->lruq);
2754 if (tdc->lruq.prev == &tdc->lruq)
2755 osi_Panic("lruq 3");
2757 /* initialize entry */
2758 tdc->f.fid.Cell = 0;
2759 tdc->f.fid.Fid.Volume = 0;
2761 hones(tdc->f.versionNo);
2762 tdc->f.inode.mem = aslot;
2763 tdc->dflags |= DFEntryMod;
2766 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2769 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2770 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2771 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2774 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2775 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2776 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2777 ObtainReadLock(&tdc->tlock);
2779 afs_indexTable[aslot] = tdc;
2782 } /*afs_MemGetDSlot */
2784 unsigned int last_error = 0, lasterrtime = 0;
2790 * Return a pointer to an freshly initialized dcache entry using
2791 * a UFS-based disk cache. The dcache tlock will be read-locked.
2794 * aslot : Dcache slot to look at.
2795 * indexvalid : 1 if we know the slot we're giving is valid, and thus
2796 * reading the dcache from the disk index should succeed. 0
2797 * if we are initializing a new dcache, and so reading from
2798 * the disk index may fail.
2799 * datavalid : 0 if we are loading a dcache entry from the free or
2800 * discard list, so we know the data in the given dcache is
2801 * not valid. 1 if we are loading a known used dcache, so the
2802 * data in the dcache must be valid.
2805 * afs_xdcache lock write-locked.
2808 afs_UFSGetDSlot(afs_int32 aslot, int indexvalid, int datavalid)
2816 AFS_STATCNT(afs_UFSGetDSlot);
2817 if (CheckLock(&afs_xdcache) != -1)
2818 osi_Panic("getdslot nolock");
2819 if (aslot < 0 || aslot >= afs_cacheFiles)
2820 osi_Panic("getdslot slot %d (of %d)", aslot, afs_cacheFiles);
2821 tdc = afs_indexTable[aslot];
2823 QRemove(&tdc->lruq); /* move to queue head */
2824 QAdd(&afs_DLRU, &tdc->lruq);
2825 /* Grab tlock in case refCount != 0 */
2826 ObtainWriteLock(&tdc->tlock, 625);
2828 ConvertWToRLock(&tdc->tlock);
2832 /* otherwise we should read it in from the cache file */
2833 if (!afs_freeDSList)
2834 afs_GetDownDSlot(4);
2835 if (!afs_freeDSList) {
2836 /* none free, making one is better than a panic */
2837 afs_stats_cmperf.dcacheXAllocs++; /* count in case we have a leak */
2838 tdc = afs_osi_Alloc(sizeof(struct dcache));
2839 osi_Assert(tdc != NULL);
2840 #ifdef KERNEL_HAVE_PIN
2841 pin((char *)tdc, sizeof(struct dcache)); /* XXX */
2844 tdc = afs_freeDSList;
2845 afs_freeDSList = (struct dcache *)tdc->lruq.next;
2848 tdc->dflags = 0; /* up-to-date, not in free q */
2850 QAdd(&afs_DLRU, &tdc->lruq);
2851 if (tdc->lruq.prev == &tdc->lruq)
2852 osi_Panic("lruq 3");
2855 * Seek to the aslot'th entry and read it in.
2857 off = sizeof(struct fcache)*aslot + sizeof(struct afs_fheader);
2859 afs_osi_Read(afs_cacheInodep,
2860 off, (char *)(&tdc->f),
2861 sizeof(struct fcache));
2863 if (code != sizeof(struct fcache)) {
2865 #if defined(KERNEL_HAVE_UERROR)
2866 last_error = getuerror();
2870 lasterrtime = osi_Time();
2872 struct osi_stat tstat;
2873 if (afs_osi_Stat(afs_cacheInodep, &tstat)) {
2876 afs_warn("afs: disk cache read error in CacheItems slot %d "
2877 "off %d/%d code %d/%d\n",
2879 off, (int)tstat.size,
2880 (int)code, (int)sizeof(struct fcache));
2881 /* put tdc back on the free dslot list */
2882 QRemove(&tdc->lruq);
2883 tdc->index = NULLIDX;
2884 tdc->lruq.next = (struct afs_q *)afs_freeDSList;
2885 afs_freeDSList = tdc;
2889 if (!afs_CellNumValid(tdc->f.fid.Cell)) {
2892 osi_Panic("afs: needed valid dcache but index %d off %d has "
2893 "invalid cell num %d\n",
2894 (int)aslot, off, (int)tdc->f.fid.Cell);
2898 if (datavalid && tdc->f.fid.Fid.Volume == 0) {
2899 osi_Panic("afs: invalid zero-volume dcache entry at slot %d off %d",
2903 if (indexvalid && !datavalid) {
2904 /* we know that the given dslot does exist, but the data in it is not
2905 * valid. this only occurs when we pull a dslot from the free or
2906 * discard list, so be sure not to re-use the data; force invalidation.
2912 tdc->f.fid.Cell = 0;
2913 tdc->f.fid.Fid.Volume = 0;
2915 hones(tdc->f.versionNo);
2916 tdc->dflags |= DFEntryMod;
2917 afs_indexUnique[aslot] = tdc->f.fid.Fid.Unique;
2918 tdc->f.states &= ~(DRO|DBackup|DRW);
2919 afs_DCMoveBucket(tdc, 0, 0);
2922 if (tdc->f.states & DRO) {
2923 afs_DCMoveBucket(tdc, 0, 2);
2924 } else if (tdc->f.states & DBackup) {
2925 afs_DCMoveBucket(tdc, 0, 1);
2927 afs_DCMoveBucket(tdc, 0, 1);
2933 if (tdc->f.chunk >= 0)
2934 tdc->validPos = AFS_CHUNKTOBASE(tdc->f.chunk) + tdc->f.chunkBytes;
2939 osi_Assert(0 == NBObtainWriteLock(&tdc->lock, 674));
2940 osi_Assert(0 == NBObtainWriteLock(&tdc->mflock, 675));
2941 osi_Assert(0 == NBObtainWriteLock(&tdc->tlock, 676));
2944 AFS_RWLOCK_INIT(&tdc->lock, "dcache lock");
2945 AFS_RWLOCK_INIT(&tdc->tlock, "dcache tlock");
2946 AFS_RWLOCK_INIT(&tdc->mflock, "dcache flock");
2947 ObtainReadLock(&tdc->tlock);
2950 * If we didn't read into a temporary dcache region, update the
2951 * slot pointer table.
2953 afs_indexTable[aslot] = tdc;
2956 } /*afs_UFSGetDSlot */
2961 * Write a particular dcache entry back to its home in the
2964 * \param adc Pointer to the dcache entry to write.
2965 * \param atime If true, set the modtime on the file to the current time.
2967 * \note Environment:
2968 * Must be called with the afs_xdcache lock at least read-locked,
2969 * and dcache entry at least read-locked.
2970 * The reference count is not changed.
2974 afs_WriteDCache(struct dcache *adc, int atime)
2978 if (cacheDiskType == AFS_FCACHE_TYPE_MEM)
2980 AFS_STATCNT(afs_WriteDCache);
2981 osi_Assert(WriteLocked(&afs_xdcache));
2983 adc->f.modTime = osi_Time();
2985 if ((afs_indexFlags[adc->index] & (IFFree | IFDiscarded)) == 0 &&
2986 adc->f.fid.Fid.Volume == 0) {
2987 /* If a dcache slot is not on the free or discard list, it must be
2988 * in the hash table. Thus, the volume must be non-zero, since that
2989 * is how we determine whether or not to unhash the entry when kicking
2990 * it out of the cache. Do this check now, since otherwise this can
2991 * cause hash table corruption and a panic later on after we read the
2993 osi_Panic("afs_WriteDCache zero volume index %d flags 0x%x\n",
2994 adc->index, (unsigned)afs_indexFlags[adc->index]);
2998 * Seek to the right dcache slot and write the in-memory image out to disk.
3000 afs_cellname_write();
3002 afs_osi_Write(afs_cacheInodep,
3003 sizeof(struct fcache) * adc->index +
3004 sizeof(struct afs_fheader), (char *)(&adc->f),
3005 sizeof(struct fcache));
3006 if (code != sizeof(struct fcache)) {
3007 afs_warn("afs: failed to write to CacheItems off %ld code %d/%d\n",
3008 (long)(sizeof(struct fcache) * adc->index + sizeof(struct afs_fheader)),
3009 (int)code, (int)sizeof(struct fcache));
3018 * Wake up users of a particular file waiting for stores to take
3021 * \param avc Ptr to related vcache entry.
3023 * \note Environment:
3024 * Nothing interesting.
3027 afs_wakeup(struct vcache *avc)
3030 struct brequest *tb;
3032 AFS_STATCNT(afs_wakeup);
3033 for (i = 0; i < NBRS; i++, tb++) {
3034 /* if request is valid and for this file, we've found it */
3035 if (tb->refCount > 0 && avc == tb->vc) {
3038 * If CSafeStore is on, then we don't awaken the guy
3039 * waiting for the store until the whole store has finished.
3040 * Otherwise, we do it now. Note that if CSafeStore is on,
3041 * the BStore routine actually wakes up the user, instead
3043 * I think this is redundant now because this sort of thing
3044 * is already being handled by the higher-level code.
3046 if ((avc->f.states & CSafeStore) == 0) {
3048 tb->flags |= BUVALID;
3049 if (tb->flags & BUWAIT) {
3050 tb->flags &= ~BUWAIT;
3061 * Given a file name and inode, set up that file to be an
3062 * active member in the AFS cache. This also involves checking
3063 * the usability of its data.
3065 * \param afile Name of the cache file to initialize.
3066 * \param ainode Inode of the file.
3068 * \note Environment:
3069 * This function is called only during initialization.
3072 afs_InitCacheFile(char *afile, ino_t ainode)
3077 struct osi_file *tfile;
3078 struct osi_stat tstat;
3081 AFS_STATCNT(afs_InitCacheFile);
3082 index = afs_stats_cmperf.cacheNumEntries;
3083 if (index >= afs_cacheFiles)
3086 ObtainWriteLock(&afs_xdcache, 282);
3087 tdc = afs_GetNewDSlot(index);
3088 ReleaseReadLock(&tdc->tlock);
3089 ReleaseWriteLock(&afs_xdcache);
3091 ObtainWriteLock(&tdc->lock, 621);
3092 ObtainWriteLock(&afs_xdcache, 622);
3093 if (!afile && !ainode) {
3098 code = afs_LookupInodeByPath(afile, &tdc->f.inode.ufs, NULL);
3100 ReleaseWriteLock(&afs_xdcache);
3101 ReleaseWriteLock(&tdc->lock);
3106 /* Add any other 'complex' inode types here ... */
3107 #if !defined(AFS_LINUX26_ENV) && !defined(AFS_CACHE_VNODE_PATH)
3108 tdc->f.inode.ufs = ainode;
3110 osi_Panic("Can't init cache with inode numbers when complex inodes are "
3115 if ((tdc->f.states & DWriting) || tdc->f.fid.Fid.Volume == 0)
3117 tfile = osi_UFSOpen(&tdc->f.inode);
3118 code = afs_osi_Stat(tfile, &tstat);
3120 osi_Panic("initcachefile stat");
3123 * If file size doesn't match the cache info file, it's probably bad.
3125 if (tdc->f.chunkBytes != tstat.size)
3128 * If file changed within T (120?) seconds of cache info file, it's
3129 * probably bad. In addition, if slot changed within last T seconds,
3130 * the cache info file may be incorrectly identified, and so slot
3133 if (cacheInfoModTime < tstat.mtime + 120)
3135 if (cacheInfoModTime < tdc->f.modTime + 120)
3137 /* In case write through is behind, make sure cache items entry is
3138 * at least as new as the chunk.
3140 if (tdc->f.modTime < tstat.mtime)
3143 tdc->f.chunkBytes = 0;
3146 tdc->f.fid.Fid.Volume = 0; /* not in the hash table */
3147 if (tfile && tstat.size != 0)
3148 osi_UFSTruncate(tfile, 0);
3149 tdc->f.states &= ~(DRO|DBackup|DRW);
3150 afs_DCMoveBucket(tdc, 0, 0);
3151 /* put entry in free cache slot list */
3152 afs_dvnextTbl[tdc->index] = afs_freeDCList;
3153 afs_freeDCList = index;
3155 afs_indexFlags[index] |= IFFree;
3156 afs_indexUnique[index] = 0;
3159 * We must put this entry in the appropriate hash tables.
3160 * Note that i is still set from the above DCHash call
3162 code = DCHash(&tdc->f.fid, tdc->f.chunk);
3163 afs_dcnextTbl[tdc->index] = afs_dchashTbl[code];
3164 afs_dchashTbl[code] = tdc->index;
3165 code = DVHash(&tdc->f.fid);
3166 afs_dvnextTbl[tdc->index] = afs_dvhashTbl[code];
3167 afs_dvhashTbl[code] = tdc->index;
3168 afs_AdjustSize(tdc, tstat.size); /* adjust to new size */
3170 /* has nontrivial amt of data */
3171 afs_indexFlags[index] |= IFEverUsed;
3172 afs_stats_cmperf.cacheFilesReused++;
3174 * Initialize index times to file's mod times; init indexCounter
3177 hset32(afs_indexTimes[index], tstat.atime);
3178 if (hgetlo(afs_indexCounter) < tstat.atime) {
3179 hset32(afs_indexCounter, tstat.atime);
3181 afs_indexUnique[index] = tdc->f.fid.Fid.Unique;
3182 } /*File is not bad */
3185 osi_UFSClose(tfile);
3186 tdc->f.states &= ~DWriting;
3187 tdc->dflags &= ~DFEntryMod;
3188 /* don't set f.modTime; we're just cleaning up */
3189 osi_Assert(afs_WriteDCache(tdc, 0) == 0);
3190 ReleaseWriteLock(&afs_xdcache);
3191 ReleaseWriteLock(&tdc->lock);
3193 afs_stats_cmperf.cacheNumEntries++;
3198 /*Max # of struct dcache's resident at any time*/
3200 * If 'dchint' is enabled then in-memory dcache min is increased because of
3206 * Initialize dcache related variables.
3216 afs_dcacheInit(int afiles, int ablocks, int aDentries, int achunk, int aflags)
3222 afs_freeDCList = NULLIDX;
3223 afs_discardDCList = NULLIDX;
3224 afs_freeDCCount = 0;
3225 afs_freeDSList = NULL;
3226 hzero(afs_indexCounter);
3228 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3234 if (achunk < 0 || achunk > 30)
3235 achunk = 13; /* Use default */
3236 AFS_SETCHUNKSIZE(achunk);
3242 if (aDentries > 512)
3243 afs_dhashsize = 2048;
3244 /* initialize hash tables */
3245 afs_dvhashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3246 osi_Assert(afs_dvhashTbl != NULL);
3247 afs_dchashTbl = afs_osi_Alloc(afs_dhashsize * sizeof(afs_int32));
3248 osi_Assert(afs_dchashTbl != NULL);
3249 for (i = 0; i < afs_dhashsize; i++) {
3250 afs_dvhashTbl[i] = NULLIDX;
3251 afs_dchashTbl[i] = NULLIDX;
3253 afs_dvnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3254 osi_Assert(afs_dvnextTbl != NULL);
3255 afs_dcnextTbl = afs_osi_Alloc(afiles * sizeof(afs_int32));
3256 osi_Assert(afs_dcnextTbl != NULL);
3257 for (i = 0; i < afiles; i++) {
3258 afs_dvnextTbl[i] = NULLIDX;
3259 afs_dcnextTbl[i] = NULLIDX;
3262 /* Allocate and zero the pointer array to the dcache entries */
3263 afs_indexTable = afs_osi_Alloc(sizeof(struct dcache *) * afiles);
3264 osi_Assert(afs_indexTable != NULL);
3265 memset(afs_indexTable, 0, sizeof(struct dcache *) * afiles);
3266 afs_indexTimes = afs_osi_Alloc(afiles * sizeof(afs_hyper_t));
3267 osi_Assert(afs_indexTimes != NULL);
3268 memset(afs_indexTimes, 0, afiles * sizeof(afs_hyper_t));
3269 afs_indexUnique = afs_osi_Alloc(afiles * sizeof(afs_uint32));
3270 osi_Assert(afs_indexUnique != NULL);
3271 memset(afs_indexUnique, 0, afiles * sizeof(afs_uint32));
3272 afs_indexFlags = afs_osi_Alloc(afiles * sizeof(u_char));
3273 osi_Assert(afs_indexFlags != NULL);
3274 memset(afs_indexFlags, 0, afiles * sizeof(char));
3276 /* Allocate and thread the struct dcache entries themselves */
3277 tdp = afs_Initial_freeDSList =
3278 afs_osi_Alloc(aDentries * sizeof(struct dcache));
3279 osi_Assert(tdp != NULL);
3280 memset(tdp, 0, aDentries * sizeof(struct dcache));
3281 #ifdef KERNEL_HAVE_PIN
3282 pin((char *)afs_indexTable, sizeof(struct dcache *) * afiles); /* XXX */
3283 pin((char *)afs_indexTimes, sizeof(afs_hyper_t) * afiles); /* XXX */
3284 pin((char *)afs_indexFlags, sizeof(char) * afiles); /* XXX */
3285 pin((char *)afs_indexUnique, sizeof(afs_int32) * afiles); /* XXX */
3286 pin((char *)tdp, aDentries * sizeof(struct dcache)); /* XXX */
3287 pin((char *)afs_dvhashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3288 pin((char *)afs_dchashTbl, sizeof(afs_int32) * afs_dhashsize); /* XXX */
3289 pin((char *)afs_dcnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3290 pin((char *)afs_dvnextTbl, sizeof(afs_int32) * afiles); /* XXX */
3293 afs_freeDSList = &tdp[0];
3294 for (i = 0; i < aDentries - 1; i++) {
3295 tdp[i].lruq.next = (struct afs_q *)(&tdp[i + 1]);
3296 AFS_RWLOCK_INIT(&tdp[i].lock, "dcache lock");
3297 AFS_RWLOCK_INIT(&tdp[i].tlock, "dcache tlock");
3298 AFS_RWLOCK_INIT(&tdp[i].mflock, "dcache flock");
3300 tdp[aDentries - 1].lruq.next = (struct afs_q *)0;
3301 AFS_RWLOCK_INIT(&tdp[aDentries - 1].lock, "dcache lock");
3302 AFS_RWLOCK_INIT(&tdp[aDentries - 1].tlock, "dcache tlock");
3303 AFS_RWLOCK_INIT(&tdp[aDentries - 1].mflock, "dcache flock");
3305 afs_stats_cmperf.cacheBlocksOrig = afs_stats_cmperf.cacheBlocksTotal =
3306 afs_cacheBlocks = ablocks;
3307 afs_ComputeCacheParms(); /* compute parms based on cache size */
3309 afs_dcentries = aDentries;
3311 afs_stats_cmperf.cacheBucket0_Discarded =
3312 afs_stats_cmperf.cacheBucket1_Discarded =
3313 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3317 if (aflags & AFSCALL_INIT_MEMCACHE) {
3319 * Use a memory cache instead of a disk cache
3321 cacheDiskType = AFS_FCACHE_TYPE_MEM;
3322 afs_cacheType = &afs_MemCacheOps;
3323 afiles = (afiles < aDentries) ? afiles : aDentries; /* min */
3324 ablocks = afiles * (AFS_FIRSTCSIZE / 1024);
3325 /* ablocks is reported in 1K blocks */
3326 code = afs_InitMemCache(afiles, AFS_FIRSTCSIZE, aflags);
3328 afs_warn("afsd: memory cache too large for available memory.\n");
3329 afs_warn("afsd: AFS files cannot be accessed.\n\n");
3332 afs_warn("Memory cache: Allocating %d dcache entries...",
3335 cacheDiskType = AFS_FCACHE_TYPE_UFS;
3336 afs_cacheType = &afs_UfsCacheOps;
3341 * Shuts down the cache.
3345 shutdown_dcache(void)
3349 #ifdef AFS_CACHE_VNODE_PATH
3350 if (cacheDiskType != AFS_FCACHE_TYPE_MEM) {
3352 for (i = 0; i < afs_cacheFiles; i++) {
3353 tdc = afs_indexTable[i];
3355 afs_osi_FreeStr(tdc->f.inode.ufs);
3361 afs_osi_Free(afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3362 afs_osi_Free(afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3363 afs_osi_Free(afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3364 afs_osi_Free(afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3365 afs_osi_Free(afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3366 afs_osi_Free(afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3367 afs_osi_Free(afs_Initial_freeDSList,
3368 afs_dcentries * sizeof(struct dcache));
3369 #ifdef KERNEL_HAVE_PIN
3370 unpin((char *)afs_dcnextTbl, afs_cacheFiles * sizeof(afs_int32));
3371 unpin((char *)afs_dvnextTbl, afs_cacheFiles * sizeof(afs_int32));
3372 unpin((char *)afs_indexTable, afs_cacheFiles * sizeof(struct dcache *));
3373 unpin((char *)afs_indexTimes, afs_cacheFiles * sizeof(afs_hyper_t));
3374 unpin((char *)afs_indexUnique, afs_cacheFiles * sizeof(afs_uint32));
3375 unpin((u_char *) afs_indexFlags, afs_cacheFiles * sizeof(u_char));
3376 unpin(afs_Initial_freeDSList, afs_dcentries * sizeof(struct dcache));
3380 for (i = 0; i < afs_dhashsize; i++) {
3381 afs_dvhashTbl[i] = NULLIDX;
3382 afs_dchashTbl[i] = NULLIDX;
3385 afs_osi_Free(afs_dvhashTbl, afs_dhashsize * sizeof(afs_int32));
3386 afs_osi_Free(afs_dchashTbl, afs_dhashsize * sizeof(afs_int32));
3388 afs_blocksUsed = afs_dcentries = 0;
3389 afs_stats_cmperf.cacheBucket0_Discarded =
3390 afs_stats_cmperf.cacheBucket1_Discarded =
3391 afs_stats_cmperf.cacheBucket2_Discarded = 0;
3392 hzero(afs_indexCounter);
3394 afs_freeDCCount = 0;
3395 afs_freeDCList = NULLIDX;
3396 afs_discardDCList = NULLIDX;
3397 afs_freeDSList = afs_Initial_freeDSList = 0;
3399 LOCK_INIT(&afs_xdcache, "afs_xdcache");
3405 * Get a dcache ready for writing, respecting the current cache size limits
3407 * len is required because afs_GetDCache with flag == 4 expects the length
3408 * field to be filled. It decides from this whether it's necessary to fetch
3409 * data into the chunk before writing or not (when the whole chunk is
3412 * \param avc The vcache to fetch a dcache for
3413 * \param filePos The start of the section to be written
3414 * \param len The length of the section to be written
3418 * \return If successful, a reference counted dcache with tdc->lock held. Lock
3419 * must be released and afs_PutDCache() called to free dcache.
3422 * \note avc->lock must be held on entry. Function may release and reobtain
3423 * avc->lock and GLOCK.
3427 afs_ObtainDCacheForWriting(struct vcache *avc, afs_size_t filePos,
3428 afs_size_t len, struct vrequest *areq,
3431 struct dcache *tdc = NULL;
3434 /* read the cached info */
3436 tdc = afs_FindDCache(avc, filePos);
3438 ObtainWriteLock(&tdc->lock, 657);
3439 } else if (afs_blocksUsed >
3440 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3441 tdc = afs_FindDCache(avc, filePos);
3443 ObtainWriteLock(&tdc->lock, 658);
3444 if (!hsame(tdc->f.versionNo, avc->f.m.DataVersion)
3445 || (tdc->dflags & DFFetching)) {
3446 ReleaseWriteLock(&tdc->lock);
3452 afs_MaybeWakeupTruncateDaemon();
3453 while (afs_blocksUsed >
3454 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3455 ReleaseWriteLock(&avc->lock);
3456 if (afs_blocksUsed - afs_blocksDiscarded >
3457 PERCENT(CM_WAITFORDRAINPCT, afs_cacheBlocks)) {
3458 afs_WaitForCacheDrain = 1;
3459 afs_osi_Sleep(&afs_WaitForCacheDrain);
3461 afs_MaybeFreeDiscardedDCache();
3462 afs_MaybeWakeupTruncateDaemon();
3463 ObtainWriteLock(&avc->lock, 509);
3465 avc->f.states |= CDirty;
3466 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3468 ObtainWriteLock(&tdc->lock, 659);
3471 tdc = afs_GetDCache(avc, filePos, areq, &offset, &len, 4);
3473 ObtainWriteLock(&tdc->lock, 660);
3476 if (!(afs_indexFlags[tdc->index] & IFDataMod)) {
3477 afs_stats_cmperf.cacheCurrDirtyChunks++;
3478 afs_indexFlags[tdc->index] |= IFDataMod; /* so it doesn't disappear */
3480 if (!(tdc->f.states & DWriting)) {
3481 /* don't mark entry as mod if we don't have to */
3482 tdc->f.states |= DWriting;
3483 tdc->dflags |= DFEntryMod;
3490 * Make a shadow copy of a dir's dcache. It's used for disconnected
3491 * operations like remove/create/rename to keep the original directory data.
3492 * On reconnection, we can diff the original data with the server and get the
3493 * server changes and with the local data to get the local changes.
3495 * \param avc The dir vnode.
3496 * \param adc The dir dcache.
3498 * \return 0 for success.
3500 * \note The vcache entry must be write locked.
3501 * \note The dcache entry must be read locked.
3504 afs_MakeShadowDir(struct vcache *avc, struct dcache *adc)
3506 int i, code, ret_code = 0, written, trans_size;
3507 struct dcache *new_dc = NULL;
3508 struct osi_file *tfile_src, *tfile_dst;
3509 struct VenusFid shadow_fid;
3512 /* Is this a dir? */
3513 if (vType(avc) != VDIR)
3516 if (avc->f.shadow.vnode || avc->f.shadow.unique)
3519 /* Generate a fid for the shadow dir. */
3520 shadow_fid.Cell = avc->f.fid.Cell;
3521 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3522 afs_GenShadowFid(&shadow_fid);
3524 ObtainWriteLock(&afs_xdcache, 716);
3526 /* Get a fresh dcache. */
3527 new_dc = afs_AllocDCache(avc, 0, 0, &shadow_fid);
3530 ObtainReadLock(&adc->mflock);
3532 /* Set up the new fid. */
3533 /* Copy interesting data from original dir dcache. */
3534 new_dc->mflags = adc->mflags;
3535 new_dc->dflags = adc->dflags;
3536 new_dc->f.modTime = adc->f.modTime;
3537 new_dc->f.versionNo = adc->f.versionNo;
3538 new_dc->f.states = adc->f.states;
3539 new_dc->f.chunk= adc->f.chunk;
3540 new_dc->f.chunkBytes = adc->f.chunkBytes;
3542 ReleaseReadLock(&adc->mflock);
3544 /* Now add to the two hash chains */
3545 i = DCHash(&shadow_fid, 0);
3546 afs_dcnextTbl[new_dc->index] = afs_dchashTbl[i];
3547 afs_dchashTbl[i] = new_dc->index;
3549 i = DVHash(&shadow_fid);
3550 afs_dvnextTbl[new_dc->index] = afs_dvhashTbl[i];
3551 afs_dvhashTbl[i] = new_dc->index;
3553 ReleaseWriteLock(&afs_xdcache);
3555 /* Alloc a 4k block. */
3556 data = afs_osi_Alloc(4096);
3558 afs_warn("afs_MakeShadowDir: could not alloc data\n");
3563 /* Open the files. */
3564 tfile_src = afs_CFileOpen(&adc->f.inode);
3565 tfile_dst = afs_CFileOpen(&new_dc->f.inode);
3567 /* And now copy dir dcache data into this dcache,
3571 while (written < adc->f.chunkBytes) {
3572 trans_size = adc->f.chunkBytes - written;
3573 if (trans_size > 4096)
3576 /* Read a chunk from the dcache. */
3577 code = afs_CFileRead(tfile_src, written, data, trans_size);
3578 if (code < trans_size) {
3583 /* Write it to the new dcache. */
3584 code = afs_CFileWrite(tfile_dst, written, data, trans_size);
3585 if (code < trans_size) {
3590 written+=trans_size;
3593 afs_CFileClose(tfile_dst);
3594 afs_CFileClose(tfile_src);
3596 afs_osi_Free(data, 4096);
3598 ReleaseWriteLock(&new_dc->lock);
3599 afs_PutDCache(new_dc);
3602 ObtainWriteLock(&afs_xvcache, 763);
3603 ObtainWriteLock(&afs_disconDirtyLock, 765);
3604 QAdd(&afs_disconShadow, &avc->shadowq);
3605 osi_Assert((afs_RefVCache(avc) == 0));
3606 ReleaseWriteLock(&afs_disconDirtyLock);
3607 ReleaseWriteLock(&afs_xvcache);
3609 avc->f.shadow.vnode = shadow_fid.Fid.Vnode;
3610 avc->f.shadow.unique = shadow_fid.Fid.Unique;
3618 * Delete the dcaches of a shadow dir.
3620 * \param avc The vcache containing the shadow fid.
3622 * \note avc must be write locked.
3625 afs_DeleteShadowDir(struct vcache *avc)
3628 struct VenusFid shadow_fid;
3630 shadow_fid.Cell = avc->f.fid.Cell;
3631 shadow_fid.Fid.Volume = avc->f.fid.Fid.Volume;
3632 shadow_fid.Fid.Vnode = avc->f.shadow.vnode;
3633 shadow_fid.Fid.Unique = avc->f.shadow.unique;
3635 tdc = afs_FindDCacheByFid(&shadow_fid);
3637 afs_HashOutDCache(tdc, 1);
3638 afs_DiscardDCache(tdc);
3641 avc->f.shadow.vnode = avc->f.shadow.unique = 0;
3642 ObtainWriteLock(&afs_disconDirtyLock, 708);
3643 QRemove(&avc->shadowq);
3644 ReleaseWriteLock(&afs_disconDirtyLock);
3645 afs_PutVCache(avc); /* Because we held it when we added to the queue */
3649 * Populate a dcache with empty chunks up to a given file size,
3650 * used before extending a file in order to avoid 'holes' which
3651 * we can't access in disconnected mode.
3653 * \param avc The vcache which is being extended (locked)
3654 * \param alen The new length of the file
3658 afs_PopulateDCache(struct vcache *avc, afs_size_t apos, struct vrequest *areq)
3661 afs_size_t len, offset;
3662 afs_int32 start, end;
3664 /* We're doing this to deal with the situation where we extend
3665 * by writing after lseek()ing past the end of the file . If that
3666 * extension skips chunks, then those chunks won't be created, and
3667 * GetDCache will assume that they have to be fetched from the server.
3668 * So, for each chunk between the current file position, and the new
3669 * length we GetDCache for that chunk.
3672 if (AFS_CHUNK(apos) == 0 || apos <= avc->f.m.Length)
3675 if (avc->f.m.Length == 0)
3678 start = AFS_CHUNK(avc->f.m.Length)+1;
3680 end = AFS_CHUNK(apos);
3683 len = AFS_CHUNKTOSIZE(start);
3684 tdc = afs_GetDCache(avc, AFS_CHUNKTOBASE(start), areq, &offset, &len, 4);